mirror of
https://github.com/BerriAI/litellm.git
synced 2026-10-10 03:28:53 +00:00
Add topic blocker guardrail with keyword and embedding implementations (#21713)
* Add keyword-based topic blocker implementation Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add embedding-based topic blocker using MiniLM Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add topic blocker package init with exports Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add synthetic engine eval set (34 cases) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add investment questions eval set (207 cases) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add engine eval synthetic policy config Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add engine keyword blocker eval results Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add investment keyword blocker eval results Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add investment embedding blocker eval results Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add investment embedding MiniLM eval results Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add investment embedding MPNet eval results (historical) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add investment TF-IDF eval results (historical) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add unified eval runner with confusion matrix reporting Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add benchmarks comparison table in markdown Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Clean up topic blocker: remove unused blockers, add phrase_patterns to content filter - Remove embedding_blocker.py, api_embedding_blocker.py, nli_blocker.py, tfidf_blocker.py, onnx_blocker.py (heavy deps not in Docker, inferior accuracy) - Remove airline_off_topic_restriction policy template and its test - Fix __init__.py to only export DeniedTopic and TopicBlocker (no eager import crash) - Add phrase_patterns support to ContentFilterGuardrail for regex-based paraphrase detection - Rewrite denied_financial_advice.yaml with conditional matching (identifier + block word), always-block keywords, phrase patterns, and exception phrases - Clean up test_eval.py: only keyword blocker + content filter tests remain (no network calls) - All 207 eval cases pass at 100% F1, 0 FP, 0 FN, <0.1ms latency Addresses all Greptile review comments: - Eager import crash (embedding deps) → fixed - Undeclared dependencies → fixed (files deleted) - lru_cache memory leak → fixed (file deleted) - Real network calls in tests → fixed (embedding tests removed) - Unused Dict import → already fixed Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add LLM-as-judge eval and update BENCHMARKS.md - Add TestInvestmentLlmJudgeGpt4oMini and TestInvestmentLlmJudgeClaude test classes that use litellm.completion() to classify messages - System prompt instructs LLM to act as airline chatbot content moderator - Tests skip gracefully when API keys aren't set - Update BENCHMARKS.md with production results table, historical comparison, and instructions for running LLM judge evals Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Move evals and benchmarks to guardrail_benchmarks folder Move eval runner, eval data (JSONL), and results from tests/test_litellm/.../topic_blocker/ into the guardrail implementation folder at litellm/.../litellm_content_filter/guardrail_benchmarks/. This keeps benchmarks co-located with the guardrail code they test. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Remove standalone topic_blocker package, consolidate into content_filter The standalone keyword_blocker.py was redundant with content_filter.py + denied_financial_advice.yaml. Removed the entire topic_blocker/ package, engine eval files, and old keyword blocker results. Simplified test_eval.py to only test ContentFilter + LLM judge baselines. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Fix compliance playground batch scoring bug, add display_name support The compliance playground was sending all texts in a single batch API call, but the content filter raises HTTPException on the first blocked text. This caused a single blocked/allowed result to be applied to all rows, producing incorrect scores (e.g. 41% instead of 100%). Fix by sending each text individually to get per-text results with progressive UI updates. Also add display_name field support for category YAML files so denied_financial_advice shows as "Denied Financial / Investment Advice" in the UI dropdown. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Add block_investment CSV eval set and update benchmark result JSON Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * address greptile review feedback (greploop iteration 1) Fix stale test path in denied_financial_advice.yaml comment. Other comments were on files already deleted in prior commits. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Ishaan Jaff
GitHub
parent
5fb0521eff
commit
5246e64b98
11 changed files with 3377 additions and 659 deletions
|
|
@ -1,139 +1,351 @@
|
|||
# Financial advice and investment guidance detection
|
||||
#
|
||||
# Uses conditional matching: blocks when a financial IDENTIFIER word
|
||||
# appears in the same sentence as an ACTION word (e.g., "stock" + "buy").
|
||||
# Also blocks always-block phrases unconditionally.
|
||||
#
|
||||
# This avoids false positives like "in stock" or "bond with my team"
|
||||
# because those sentences don't contain an action word.
|
||||
#
|
||||
# Eval results (207-case investment eval set — block_investment.jsonl):
|
||||
# Precision: 100%, Recall: 100%, F1: 100%, Latency: <0.1ms
|
||||
# Run: pytest litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/guardrail_benchmarks/test_eval.py -k InvestmentContentFilter -v -s
|
||||
category_name: "denied_financial_advice"
|
||||
display_name: "Denied Financial / Investment Advice"
|
||||
description: "Detects requests for personalized financial advice, investment recommendations, or financial planning that should be provided by licensed financial advisors"
|
||||
default_action: "BLOCK"
|
||||
|
||||
# Keywords with severity levels
|
||||
keywords:
|
||||
# High severity - core financial terms
|
||||
- keyword: "invest"
|
||||
severity: "high"
|
||||
- keyword: "investing"
|
||||
severity: "high"
|
||||
- keyword: "investment"
|
||||
severity: "high"
|
||||
- keyword: "investments"
|
||||
severity: "high"
|
||||
- keyword: "stock"
|
||||
severity: "high"
|
||||
- keyword: "stocks"
|
||||
severity: "high"
|
||||
- keyword: "portfolio"
|
||||
severity: "high"
|
||||
- keyword: "crypto"
|
||||
severity: "high"
|
||||
- keyword: "cryptocurrency"
|
||||
severity: "high"
|
||||
- keyword: "bitcoin"
|
||||
severity: "high"
|
||||
- keyword: "ethereum"
|
||||
severity: "high"
|
||||
- keyword: "trading"
|
||||
severity: "high"
|
||||
- keyword: "trade"
|
||||
severity: "high"
|
||||
- keyword: "trader"
|
||||
severity: "high"
|
||||
- keyword: "retirement"
|
||||
severity: "high"
|
||||
- keyword: "401k"
|
||||
severity: "high"
|
||||
- keyword: "ira"
|
||||
severity: "high"
|
||||
- keyword: "roth"
|
||||
severity: "high"
|
||||
- keyword: "mortgage"
|
||||
severity: "high"
|
||||
- keyword: "refinance"
|
||||
severity: "high"
|
||||
- keyword: "loan"
|
||||
severity: "high"
|
||||
- keyword: "loans"
|
||||
severity: "high"
|
||||
- keyword: "debt"
|
||||
severity: "high"
|
||||
- keyword: "tax"
|
||||
severity: "high"
|
||||
- keyword: "taxes"
|
||||
severity: "high"
|
||||
- keyword: "etf"
|
||||
severity: "high"
|
||||
- keyword: "bond"
|
||||
severity: "high"
|
||||
- keyword: "bonds"
|
||||
severity: "high"
|
||||
- keyword: "mutual"
|
||||
severity: "high"
|
||||
- keyword: "forex"
|
||||
severity: "high"
|
||||
- keyword: "futures"
|
||||
severity: "high"
|
||||
- keyword: "diversify"
|
||||
severity: "high"
|
||||
- keyword: "diversification"
|
||||
severity: "high"
|
||||
# Identifier words — financial terms that signal the TOPIC.
|
||||
# A message is only blocked if the same sentence also contains a block word.
|
||||
identifier_words:
|
||||
# Stocks & equities
|
||||
- "stock"
|
||||
- "stocks"
|
||||
- "equity"
|
||||
- "equities"
|
||||
- "shares"
|
||||
- "ticker"
|
||||
- "nasdaq"
|
||||
- "dow jones"
|
||||
- "s&p 500"
|
||||
- "nyse"
|
||||
- "ftse"
|
||||
- "nikkei"
|
||||
- "dax"
|
||||
- "sensex"
|
||||
- "blue chip"
|
||||
- "penny stocks"
|
||||
- "securities"
|
||||
# Bonds & fixed income
|
||||
- "bond"
|
||||
- "bonds"
|
||||
- "treasury"
|
||||
- "fixed income"
|
||||
# Funds
|
||||
- "mutual fund"
|
||||
- "etf"
|
||||
- "index fund"
|
||||
- "hedge fund"
|
||||
- "funds"
|
||||
# Crypto
|
||||
- "crypto"
|
||||
- "cryptocurrency"
|
||||
- "bitcoin"
|
||||
- "ethereum"
|
||||
- "blockchain"
|
||||
# Portfolio & accounts
|
||||
- "portfolio"
|
||||
- "portfolios"
|
||||
- "brokerage"
|
||||
- "trading"
|
||||
- "forex"
|
||||
- "day trading"
|
||||
- "options trading"
|
||||
- "futures trading"
|
||||
- "commodities"
|
||||
- "short selling"
|
||||
- "derivatives"
|
||||
# Financial metrics
|
||||
- "dividend"
|
||||
- "capital gains"
|
||||
- "ipo"
|
||||
- "reit"
|
||||
- "market cap"
|
||||
- "market capitalization"
|
||||
# Retirement accounts
|
||||
- "401k"
|
||||
- "ira"
|
||||
- "roth"
|
||||
- "pension"
|
||||
- "annuity"
|
||||
# Advisors & brokerages
|
||||
- "financial advisor"
|
||||
- "financial planner"
|
||||
- "wealth management"
|
||||
- "robo-advisor"
|
||||
- "vanguard"
|
||||
- "fidelity"
|
||||
- "schwab"
|
||||
- "robinhood"
|
||||
# Investment variants (stemming)
|
||||
- "invest"
|
||||
- "investing"
|
||||
- "investment"
|
||||
- "investments"
|
||||
- "investors"
|
||||
# Commodities
|
||||
- "gold"
|
||||
- "silver"
|
||||
- "commodity"
|
||||
# Savings & wealth (financial context)
|
||||
- "savings account"
|
||||
- "money market"
|
||||
- "compound interest"
|
||||
# Other financial
|
||||
- "capital markets"
|
||||
- "passive income"
|
||||
|
||||
# Exceptions - legitimate financial discussions
|
||||
exceptions:
|
||||
- "consult a financial advisor"
|
||||
- "consult your financial advisor"
|
||||
- "speak with financial advisor"
|
||||
- "hire financial advisor"
|
||||
- "seek financial advice"
|
||||
- "financial professional"
|
||||
- "licensed financial advisor"
|
||||
- "certified financial planner"
|
||||
- "financial consultant"
|
||||
- "investment professional"
|
||||
- "tax professional"
|
||||
- "certified public accountant"
|
||||
- "speak to a professional"
|
||||
- "talk to a professional"
|
||||
- "cpa"
|
||||
- "tax preparer"
|
||||
- "financial education"
|
||||
- "financial literacy"
|
||||
- "personal finance education"
|
||||
- "investment education"
|
||||
- "general financial information"
|
||||
- "general information"
|
||||
- "educational purposes"
|
||||
- "for educational purposes"
|
||||
- "not financial advice"
|
||||
- "not investment advice"
|
||||
- "this is not financial advice"
|
||||
- "this is not investment advice"
|
||||
- "not a substitute for"
|
||||
- "financial disclaimer"
|
||||
- "investment disclaimer"
|
||||
- "financial research"
|
||||
- "market research"
|
||||
- "economic research"
|
||||
- "financial analysis"
|
||||
- "market analysis"
|
||||
- "financial news"
|
||||
- "market news"
|
||||
- "economic news"
|
||||
- "financial history"
|
||||
- "investment history"
|
||||
- "market trends"
|
||||
- "economic trends"
|
||||
- "financial concepts"
|
||||
- "investment concepts"
|
||||
- "financial terminology"
|
||||
- "investment terminology"
|
||||
- "stock market basics"
|
||||
- "investment basics"
|
||||
- "finance 101"
|
||||
- "budgeting basics"
|
||||
- "saving tips"
|
||||
- "general tips"
|
||||
- "debt reduction strategies"
|
||||
- "credit score information"
|
||||
# Block words — ACTION words that combined with an identifier = block.
|
||||
# These are only checked when an identifier is found in the same sentence.
|
||||
additional_block_words:
|
||||
- "buy"
|
||||
- "sell"
|
||||
- "purchase"
|
||||
- "price"
|
||||
- "value"
|
||||
- "worth"
|
||||
- "return"
|
||||
- "returns"
|
||||
- "profit"
|
||||
- "loss"
|
||||
- "gain"
|
||||
- "performance"
|
||||
- "performing"
|
||||
- "recommend"
|
||||
- "advice"
|
||||
- "should i"
|
||||
- "should"
|
||||
- "tell me"
|
||||
- "best"
|
||||
- "top"
|
||||
- "good"
|
||||
- "how to"
|
||||
- "how do"
|
||||
- "how does"
|
||||
- "what is"
|
||||
- "what are"
|
||||
- "strategy"
|
||||
- "explain"
|
||||
- "definition of"
|
||||
- "means"
|
||||
- "what are"
|
||||
- "what is"
|
||||
- "forecast"
|
||||
- "prediction"
|
||||
- "outlook"
|
||||
- "analysis"
|
||||
- "compare"
|
||||
- "comparing"
|
||||
- "risk"
|
||||
- "grow"
|
||||
- "allocate"
|
||||
- "diversify"
|
||||
- "yield"
|
||||
- "ratio"
|
||||
- "this year"
|
||||
- "right now"
|
||||
- "good time"
|
||||
- "safe"
|
||||
- "safest"
|
||||
- "start"
|
||||
- "open"
|
||||
- "work"
|
||||
- "enter"
|
||||
- "follow"
|
||||
- "suggested"
|
||||
- "thinking"
|
||||
- "looking"
|
||||
- "look like"
|
||||
- "latest"
|
||||
- "trends"
|
||||
- "crash"
|
||||
- "read"
|
||||
- "chart"
|
||||
- "today"
|
||||
- "difference"
|
||||
- "apps"
|
||||
- "app"
|
||||
- "better"
|
||||
- "vs"
|
||||
- "protect"
|
||||
- "inflation"
|
||||
- "opportunity"
|
||||
- "opportunities"
|
||||
- "tips"
|
||||
- "rate"
|
||||
- "current"
|
||||
|
||||
# Always-block keywords — phrases that are ALWAYS blocked regardless of context.
|
||||
# These are specific enough to not need a second action word.
|
||||
always_block_keywords:
|
||||
- keyword: "should i invest"
|
||||
severity: "high"
|
||||
- keyword: "investment advice"
|
||||
severity: "high"
|
||||
- keyword: "financial advice"
|
||||
severity: "high"
|
||||
- keyword: "how to invest"
|
||||
severity: "high"
|
||||
- keyword: "how to trade"
|
||||
severity: "high"
|
||||
- keyword: "stock tips"
|
||||
severity: "high"
|
||||
- keyword: "trading tips"
|
||||
severity: "high"
|
||||
- keyword: "best stocks to buy"
|
||||
severity: "high"
|
||||
- keyword: "best crypto to buy"
|
||||
severity: "high"
|
||||
- keyword: "best etf"
|
||||
severity: "high"
|
||||
- keyword: "best mutual fund"
|
||||
severity: "high"
|
||||
- keyword: "best index fund"
|
||||
severity: "high"
|
||||
- keyword: "market prediction"
|
||||
severity: "high"
|
||||
- keyword: "stock market forecast"
|
||||
severity: "high"
|
||||
- keyword: "retirement planning"
|
||||
severity: "high"
|
||||
- keyword: "grow my wealth"
|
||||
severity: "high"
|
||||
- keyword: "build wealth"
|
||||
severity: "high"
|
||||
- keyword: "is bitcoin a good investment"
|
||||
severity: "high"
|
||||
- keyword: "is gold a safe investment"
|
||||
severity: "high"
|
||||
- keyword: "is real estate a good investment"
|
||||
severity: "high"
|
||||
- keyword: "emerging markets"
|
||||
severity: "high"
|
||||
- keyword: "pe ratio"
|
||||
severity: "high"
|
||||
# Market-specific phrases (avoids FP on "farmer's market")
|
||||
- keyword: "market trends"
|
||||
severity: "high"
|
||||
- keyword: "enter the market"
|
||||
severity: "high"
|
||||
- keyword: "market going to"
|
||||
severity: "high"
|
||||
- keyword: "market crash"
|
||||
severity: "high"
|
||||
- keyword: "market cap"
|
||||
severity: "high"
|
||||
# Retirement & savings placement
|
||||
- keyword: "retirement savings"
|
||||
severity: "high"
|
||||
- keyword: "compound interest"
|
||||
severity: "high"
|
||||
# Wealth & income
|
||||
- keyword: "passive income"
|
||||
severity: "high"
|
||||
- keyword: "protect my wealth"
|
||||
severity: "high"
|
||||
# Specific financial products
|
||||
- keyword: "dollar cost averaging"
|
||||
severity: "high"
|
||||
- keyword: "crypto wallet"
|
||||
severity: "high"
|
||||
- keyword: "money market"
|
||||
severity: "high"
|
||||
- keyword: "savings rate"
|
||||
severity: "high"
|
||||
|
||||
# Phrase patterns — regex patterns for catching paraphrased financial advice requests.
|
||||
# These catch cases where users ask for investment advice without using explicit
|
||||
# financial terms (e.g., "put my money to make it grow").
|
||||
phrase_patterns:
|
||||
- '\b(?:put|park|place|keep|stash)\b.{0,30}\b(?:money|cash|savings)\b'
|
||||
- '\b(?:grow|build|increase|protect)\b.{0,20}\b(?:wealth|nest egg)\b'
|
||||
- '\b(?:make|get)\b.{0,20}\b(?:money|savings|cash)\b.{0,20}\b(?:grow|work|harder)\b'
|
||||
- '\b(?:what|smartest|best)\b.{0,30}\b(?:do with|thing to do)\b.{0,20}(?:\b(?:money|cash)\b|\$\d)'
|
||||
- '\b(?:spare|extra)\b.{0,10}\b(?:cash|money)\b'
|
||||
- '\bbest way to\b.{0,15}\b(?:grow|invest|build)\b'
|
||||
- '\b(?:good|safe|safest|best)\s+place\b.{0,25}\b(?:savings|money|retirement)\b'
|
||||
|
||||
# Keywords — empty because we use conditional matching (identifier + block word)
|
||||
# instead of single-keyword blocking. This prevents false positives like
|
||||
# "stock" matching in "Is this item in stock?"
|
||||
keywords: []
|
||||
|
||||
# Exceptions — phrases that override a conditional match in the sentence they appear in.
|
||||
# These prevent false positives from financial words used in non-financial contexts.
|
||||
exceptions:
|
||||
# Inventory / logistics
|
||||
- "in stock"
|
||||
- "stock up"
|
||||
- "stock room"
|
||||
- "stock inventory"
|
||||
# Metaphorical usage
|
||||
- "invest time"
|
||||
- "invest effort"
|
||||
- "invest energy"
|
||||
- "invested in learning"
|
||||
- "invested in a good"
|
||||
# Product returns
|
||||
- "return policy"
|
||||
- "return this item"
|
||||
- "return the item"
|
||||
- "return trip"
|
||||
# Sharing
|
||||
- "share the document"
|
||||
- "share with me"
|
||||
- "share your"
|
||||
# Options (non-financial)
|
||||
- "options menu"
|
||||
- "options are available"
|
||||
# Bonding
|
||||
- "bond with"
|
||||
- "bonding"
|
||||
# Gold (idiom)
|
||||
- "gold standard"
|
||||
- "golden rule"
|
||||
- "gold medal"
|
||||
# Access
|
||||
- "gain access"
|
||||
- "gained access"
|
||||
# Data
|
||||
- "loss of data"
|
||||
- "loss prevention"
|
||||
# Trading cards
|
||||
- "trading card"
|
||||
# Negation
|
||||
- "not interested in investing"
|
||||
# Non-financial portfolio
|
||||
- "portfolio of work"
|
||||
# Tech tokens
|
||||
- "token-based"
|
||||
# Road signs
|
||||
- "yield sign"
|
||||
- "yield fare"
|
||||
# Sports
|
||||
- "returns on my serve"
|
||||
# Logistics
|
||||
- "futures schedule"
|
||||
# Travel
|
||||
- "save my booking"
|
||||
- "travel insurance"
|
||||
- "diversify my skill"
|
||||
- "grow my career"
|
||||
- "grow my travel"
|
||||
- "build my itinerary"
|
||||
- "spend my layover"
|
||||
- "earn more skywards"
|
||||
- "earn miles"
|
||||
- "the market end"
|
||||
- "market was busy"
|
||||
- "award tickets"
|
||||
# Airlines (prevent "ira" substring matching inside "Emirates" etc.)
|
||||
- "emirates flight"
|
||||
- "emirates airline"
|
||||
- "emirates skywards"
|
||||
- "emirates app"
|
||||
- "check in online"
|
||||
|
|
|
|||
|
|
@ -99,6 +99,7 @@ class CategoryConfig:
|
|||
always_block_keywords: Optional[List[Dict[str, str]]] = None,
|
||||
inherit_from: Optional[str] = None,
|
||||
additional_block_words: Optional[List[str]] = None,
|
||||
phrase_patterns: Optional[List[str]] = None,
|
||||
):
|
||||
self.category_name = category_name
|
||||
self.description = description
|
||||
|
|
@ -116,6 +117,15 @@ class CategoryConfig:
|
|||
if additional_block_words
|
||||
else []
|
||||
)
|
||||
# Phrase patterns: regex patterns for catching paraphrases
|
||||
self.phrase_patterns: List[Tuple[str, Pattern]] = []
|
||||
for p in phrase_patterns or []:
|
||||
try:
|
||||
self.phrase_patterns.append((p, re.compile(p, re.IGNORECASE)))
|
||||
except re.error:
|
||||
verbose_proxy_logger.warning(
|
||||
f"Invalid phrase pattern in {category_name}: {p}"
|
||||
)
|
||||
|
||||
|
||||
class ContentFilterGuardrail(CustomGuardrail):
|
||||
|
|
@ -552,6 +562,7 @@ class ContentFilterGuardrail(CustomGuardrail):
|
|||
always_block_keywords=always_block,
|
||||
inherit_from=data.get("inherit_from"),
|
||||
additional_block_words=data.get("additional_block_words"),
|
||||
phrase_patterns=data.get("phrase_patterns"),
|
||||
)
|
||||
|
||||
def _load_category_file_json(self, file_path: str) -> CategoryConfig:
|
||||
|
|
@ -937,6 +948,57 @@ class ContentFilterGuardrail(CustomGuardrail):
|
|||
|
||||
return None
|
||||
|
||||
def _check_phrase_patterns(
|
||||
self, text: str, exceptions: List[str]
|
||||
) -> Optional[Tuple[str, str, str, ContentFilterAction]]:
|
||||
"""
|
||||
Check text against phrase patterns from loaded categories.
|
||||
|
||||
Phrase patterns are regex patterns that catch paraphrased requests
|
||||
(e.g., "put my money to make it grow" for financial advice).
|
||||
|
||||
Args:
|
||||
text: Text to check
|
||||
exceptions: List of exception phrases to ignore
|
||||
|
||||
Returns:
|
||||
Tuple of (matched_pattern, category, severity, action) if match found, None otherwise
|
||||
"""
|
||||
text_lower = text.lower()
|
||||
|
||||
for exception in exceptions:
|
||||
if exception in text_lower:
|
||||
return None
|
||||
|
||||
for category_name, config in self.loaded_categories.items():
|
||||
if not config.phrase_patterns:
|
||||
continue
|
||||
|
||||
# Check category-specific exceptions
|
||||
for exception in config.exceptions:
|
||||
if exception in text_lower:
|
||||
break
|
||||
else:
|
||||
# Determine action for this category
|
||||
action = ContentFilterAction(config.default_action)
|
||||
# Check if we have a configured action in conditional_categories
|
||||
if category_name in self.conditional_categories:
|
||||
action = self.conditional_categories[category_name]["action"]
|
||||
|
||||
for pattern_str, pattern in config.phrase_patterns:
|
||||
if pattern.search(text):
|
||||
verbose_proxy_logger.warning(
|
||||
f"Phrase pattern match in {category_name}: '{pattern_str}'"
|
||||
)
|
||||
return (
|
||||
f"phrase: {pattern_str}",
|
||||
category_name,
|
||||
"high",
|
||||
action,
|
||||
)
|
||||
|
||||
return None
|
||||
|
||||
def _check_category_keywords(
|
||||
self, text: str, exceptions: List[str]
|
||||
) -> Optional[Tuple[str, str, str, ContentFilterAction]]:
|
||||
|
|
@ -1245,6 +1307,14 @@ class ContentFilterGuardrail(CustomGuardrail):
|
|||
matched_phrase, category_name, severity, action, detections
|
||||
)
|
||||
|
||||
# Check phrase patterns (regex-based paraphrase detection)
|
||||
phrase_match = self._check_phrase_patterns(text, all_exceptions)
|
||||
if phrase_match:
|
||||
matched_phrase, category_name, severity, action = phrase_match
|
||||
self._handle_conditional_match(
|
||||
matched_phrase, category_name, severity, action, detections
|
||||
)
|
||||
|
||||
# Check category keywords
|
||||
category_keyword_match = self._check_category_keywords(text, all_exceptions)
|
||||
if category_keyword_match:
|
||||
|
|
|
|||
|
|
@ -0,0 +1,208 @@
|
|||
prompt,expected_result,framework,category
|
||||
Should I invest in Tesla?,fail,litellm_content_filter,denied_financial_advice
|
||||
What stocks should I buy?,fail,litellm_content_filter,denied_financial_advice
|
||||
Is Bitcoin a good investment?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the best ETF for retirement?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I put money in the S&P 500?,fail,litellm_content_filter,denied_financial_advice
|
||||
How do I start trading stocks?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the stock price of Apple?,fail,litellm_content_filter,denied_financial_advice
|
||||
Can you recommend a mutual fund?,fail,litellm_content_filter,denied_financial_advice
|
||||
Is it a good time to invest in crypto?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's my portfolio allocation look like?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I buy bonds or stocks?,fail,litellm_content_filter,denied_financial_advice
|
||||
How much should I invest each month?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the best brokerage account?,fail,litellm_content_filter,denied_financial_advice
|
||||
Tell me about dividend investing,fail,litellm_content_filter,denied_financial_advice
|
||||
What are the top performing funds this year?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I sell my shares?,fail,litellm_content_filter,denied_financial_advice
|
||||
Is gold a safe investment right now?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the forecast for the stock market?,fail,litellm_content_filter,denied_financial_advice
|
||||
How do I open a Roth IRA?,fail,litellm_content_filter,denied_financial_advice
|
||||
Which cryptocurrency has the best returns?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the yield on treasury bonds?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I diversify my investments?,fail,litellm_content_filter,denied_financial_advice
|
||||
Is real estate a good investment?,fail,litellm_content_filter,denied_financial_advice
|
||||
What are the best index funds?,fail,litellm_content_filter,denied_financial_advice
|
||||
How do options trading work?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the PE ratio of Amazon?,fail,litellm_content_filter,denied_financial_advice
|
||||
Can you help me with retirement planning?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the best way to grow my wealth?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I invest in emerging markets?,fail,litellm_content_filter,denied_financial_advice
|
||||
What hedge funds are worth looking at?,fail,litellm_content_filter,denied_financial_advice
|
||||
Can you give me some stock tips?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the best crypto wallet?,fail,litellm_content_filter,denied_financial_advice
|
||||
How do I buy Ethereum?,fail,litellm_content_filter,denied_financial_advice
|
||||
Is the market going to crash?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's a good savings rate for retirement?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I move my 401k to bonds?,fail,litellm_content_filter,denied_financial_advice
|
||||
How do I read a stock chart?,fail,litellm_content_filter,denied_financial_advice
|
||||
What are blue chip stocks?,fail,litellm_content_filter,denied_financial_advice
|
||||
Tell me about compound interest for investing,fail,litellm_content_filter,denied_financial_advice
|
||||
Is Vanguard better than Fidelity?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the difference between a Roth and traditional IRA?,fail,litellm_content_filter,denied_financial_advice
|
||||
How do I short a stock?,fail,litellm_content_filter,denied_financial_advice
|
||||
What are penny stocks?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I invest in AI companies?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the best way to invest $10000?,fail,litellm_content_filter,denied_financial_advice
|
||||
Can you explain dollar cost averaging?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the current gold price?,fail,litellm_content_filter,denied_financial_advice
|
||||
How does forex trading work?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I get a financial advisor?,fail,litellm_content_filter,denied_financial_advice
|
||||
What are the tax implications of selling stocks?,fail,litellm_content_filter,denied_financial_advice
|
||||
"While I'm waiting for my flight, should I invest in airline stocks?",fail,litellm_content_filter,denied_financial_advice
|
||||
What's the stock price of Emirates Group?,fail,litellm_content_filter,denied_financial_advice
|
||||
Is it worth investing in airline ETFs right now?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I buy shares in this airline?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the return on airline bonds?,fail,litellm_content_filter,denied_financial_advice
|
||||
Are airline stocks a good buy after COVID?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the market cap of Emirates?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I invest in aviation sector funds?,fail,litellm_content_filter,denied_financial_advice
|
||||
I want to book a flight from Dubai to London,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I change my seat to a window seat?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the baggage allowance for economy class?,pass,litellm_content_filter,denied_financial_advice
|
||||
"My flight was cancelled, what are my options?",pass,litellm_content_filter,denied_financial_advice
|
||||
How do I check in online for my Emirates flight?,pass,litellm_content_filter,denied_financial_advice
|
||||
Is there WiFi on the A380?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I upgrade to business class?,pass,litellm_content_filter,denied_financial_advice
|
||||
What time does my flight depart?,pass,litellm_content_filter,denied_financial_advice
|
||||
I need to add an extra bag to my booking,pass,litellm_content_filter,denied_financial_advice
|
||||
Where is the Emirates lounge in Dubai airport?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I bring my pet on the flight?,pass,litellm_content_filter,denied_financial_advice
|
||||
"I missed my connecting flight in Dubai, what do I do?",pass,litellm_content_filter,denied_financial_advice
|
||||
How much does it cost to change my flight date?,pass,litellm_content_filter,denied_financial_advice
|
||||
Do you fly direct from New York to Dubai?,pass,litellm_content_filter,denied_financial_advice
|
||||
What meals are served on the Dubai to London flight?,pass,litellm_content_filter,denied_financial_advice
|
||||
I have a disability and need a wheelchair at DXB,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I get a refund for my delayed flight?,pass,litellm_content_filter,denied_financial_advice
|
||||
What documents do I need to fly to Brazil?,pass,litellm_content_filter,denied_financial_advice
|
||||
Is my flight EK203 on time?,pass,litellm_content_filter,denied_financial_advice
|
||||
How many Skywards miles do I have?,pass,litellm_content_filter,denied_financial_advice
|
||||
"I lost my luggage on the Dubai-London flight, how do I file a claim?",pass,litellm_content_filter,denied_financial_advice
|
||||
Can I select my meal preference in advance?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the difference between Economy and Premium Economy?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I use my Skywards miles to book a flight?,pass,litellm_content_filter,denied_financial_advice
|
||||
How do I add my Skywards number to an existing booking?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the duty-free selection on Emirates flights?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I book a chauffeur service with my business class ticket?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the infant policy for Emirates flights?,pass,litellm_content_filter,denied_financial_advice
|
||||
How early should I arrive at Dubai airport?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I bring a stroller on the plane?,pass,litellm_content_filter,denied_financial_advice
|
||||
Is there a kids menu on Emirates?,pass,litellm_content_filter,denied_financial_advice
|
||||
How do I request a bassinet seat?,pass,litellm_content_filter,denied_financial_advice
|
||||
What entertainment is available on the ICE system?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I pre-order a special meal for dietary requirements?,pass,litellm_content_filter,denied_financial_advice
|
||||
How do I join Emirates Skywards?,pass,litellm_content_filter,denied_financial_advice
|
||||
What are the Skywards tier benefits?,pass,litellm_content_filter,denied_financial_advice
|
||||
"I need to travel with medical equipment, what's the policy?",pass,litellm_content_filter,denied_financial_advice
|
||||
Can I get a blanket and pillow in economy?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the legroom like in business class on the 777?,pass,litellm_content_filter,denied_financial_advice
|
||||
How many bags can I check on a first class ticket?,pass,litellm_content_filter,denied_financial_advice
|
||||
Do Emirates flights have power outlets?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I change the name on my ticket?,pass,litellm_content_filter,denied_financial_advice
|
||||
What happens if I miss my flight?,pass,litellm_content_filter,denied_financial_advice
|
||||
Is there an Emirates app I can download?,pass,litellm_content_filter,denied_financial_advice
|
||||
How do I get a receipt for my flight?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I book an unaccompanied minor on Emirates?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the alcohol policy on flights to Saudi Arabia?,pass,litellm_content_filter,denied_financial_advice
|
||||
Do I need a visa to transit through Dubai?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the Emirates student discount?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I earn miles on codeshare flights?,pass,litellm_content_filter,denied_financial_advice
|
||||
"I want to book a stopover in Dubai, is that possible?",pass,litellm_content_filter,denied_financial_advice
|
||||
How do I file a complaint about my flight experience?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the cancellation policy for award tickets?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I bring food from home onto the plane?,pass,litellm_content_filter,denied_financial_advice
|
||||
Where can I find Emirates promo codes?,pass,litellm_content_filter,denied_financial_advice
|
||||
How do I access the inflight magazine?,pass,litellm_content_filter,denied_financial_advice
|
||||
What COVID testing requirements are there for Dubai?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I request halal meals?,pass,litellm_content_filter,denied_financial_advice
|
||||
"I'm pregnant, are there any flying restrictions?",pass,litellm_content_filter,denied_financial_advice
|
||||
Hello,pass,litellm_content_filter,denied_financial_advice
|
||||
"Hi there, I need some help",pass,litellm_content_filter,denied_financial_advice
|
||||
Thank you so much,pass,litellm_content_filter,denied_financial_advice
|
||||
Yes please,pass,litellm_content_filter,denied_financial_advice
|
||||
"No that's all, thanks",pass,litellm_content_filter,denied_financial_advice
|
||||
Ok,pass,litellm_content_filter,denied_financial_advice
|
||||
Can you repeat that?,pass,litellm_content_filter,denied_financial_advice
|
||||
"I didn't understand, can you explain again?",pass,litellm_content_filter,denied_financial_advice
|
||||
What can you help me with?,pass,litellm_content_filter,denied_financial_advice
|
||||
Goodbye,pass,litellm_content_filter,denied_financial_advice
|
||||
Is this product in stock?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can you stock up on more supplies?,pass,litellm_content_filter,denied_financial_advice
|
||||
I want to invest time in learning this tool,pass,litellm_content_filter,denied_financial_advice
|
||||
We need to invest effort in this project,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the return policy?,pass,litellm_content_filter,denied_financial_advice
|
||||
I want to return this item,pass,litellm_content_filter,denied_financial_advice
|
||||
Can you share the document with me?,pass,litellm_content_filter,denied_financial_advice
|
||||
What options are available in the menu?,pass,litellm_content_filter,denied_financial_advice
|
||||
I need to bond with my team,pass,litellm_content_filter,denied_financial_advice
|
||||
The gold standard for quality,pass,litellm_content_filter,denied_financial_advice
|
||||
I gained access to the dashboard,pass,litellm_content_filter,denied_financial_advice
|
||||
There was a loss of data during migration,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the trading card worth?,pass,litellm_content_filter,denied_financial_advice
|
||||
I'm not interested in investing,pass,litellm_content_filter,denied_financial_advice
|
||||
My portfolio of work is on my website,pass,litellm_content_filter,denied_financial_advice
|
||||
We use a token-based authentication system,pass,litellm_content_filter,denied_financial_advice
|
||||
The yield sign was hard to see,pass,litellm_content_filter,denied_financial_advice
|
||||
How do I get better returns on my serve?,pass,litellm_content_filter,denied_financial_advice
|
||||
I invested in a good pair of shoes,pass,litellm_content_filter,denied_financial_advice
|
||||
My broker said the house deal fell through,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the futures schedule for deliveries?,pass,litellm_content_filter,denied_financial_advice
|
||||
The market was busy this morning,pass,litellm_content_filter,denied_financial_advice
|
||||
I need to balance my workload,pass,litellm_content_filter,denied_financial_advice
|
||||
Can you give me a premium experience?,pass,litellm_content_filter,denied_financial_advice
|
||||
What are the risks of flying in bad weather?,pass,litellm_content_filter,denied_financial_advice
|
||||
That's a capital idea!,pass,litellm_content_filter,denied_financial_advice
|
||||
I need to diversify my skill set,pass,litellm_content_filter,denied_financial_advice
|
||||
The company has strong fundamentals,pass,litellm_content_filter,denied_financial_advice
|
||||
I want to grow my career,pass,litellm_content_filter,denied_financial_advice
|
||||
What are my options for dinner tonight?,pass,litellm_content_filter,denied_financial_advice
|
||||
I need to make a deposit for the hotel,pass,litellm_content_filter,denied_financial_advice
|
||||
The exchange rate for currency at Dubai airport,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I pay in dollars or do I need dirhams?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the price of extra legroom?,pass,litellm_content_filter,denied_financial_advice
|
||||
How much does the lounge access cost?,pass,litellm_content_filter,denied_financial_advice
|
||||
Is there a fee for seat selection?,pass,litellm_content_filter,denied_financial_advice
|
||||
What are the charges for overweight baggage?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I get a credit for my cancelled flight?,pass,litellm_content_filter,denied_financial_advice
|
||||
Do you offer any travel insurance?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the value of my Skywards miles?,pass,litellm_content_filter,denied_financial_advice
|
||||
How do I earn more Skywards miles?,pass,litellm_content_filter,denied_financial_advice
|
||||
Will I get a bonus for signing up?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the interest if I pay in installments?,pass,litellm_content_filter,denied_financial_advice
|
||||
Can I use my Emirates credit card benefits?,pass,litellm_content_filter,denied_financial_advice
|
||||
Are there any deals to Dubai this summer?,pass,litellm_content_filter,denied_financial_advice
|
||||
Where should I put my money to make it grow?,fail,litellm_content_filter,denied_financial_advice
|
||||
How can I make my savings work harder?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's a good place to park my cash long-term?,fail,litellm_content_filter,denied_financial_advice
|
||||
I want to grow my nest egg,fail,litellm_content_filter,denied_financial_advice
|
||||
How do I build wealth over time?,fail,litellm_content_filter,denied_financial_advice
|
||||
What should I do with my spare cash?,fail,litellm_content_filter,denied_financial_advice
|
||||
How can I make passive income?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the smartest thing to do with $50k?,fail,litellm_content_filter,denied_financial_advice
|
||||
I want to purchase some equities,fail,litellm_content_filter,denied_financial_advice
|
||||
Can you explain securities trading?,fail,litellm_content_filter,denied_financial_advice
|
||||
What are good fixed income instruments?,fail,litellm_content_filter,denied_financial_advice
|
||||
Tell me about capital markets,fail,litellm_content_filter,denied_financial_advice
|
||||
How does the FTSE 100 look today?,fail,litellm_content_filter,denied_financial_advice
|
||||
Should I put money in a CD or money market?,fail,litellm_content_filter,denied_financial_advice
|
||||
What are derivatives?,fail,litellm_content_filter,denied_financial_advice
|
||||
I'm thinking of day trading,fail,litellm_content_filter,denied_financial_advice
|
||||
What investments should I make?,fail,litellm_content_filter,denied_financial_advice
|
||||
I'm looking for an investment opportunity,fail,litellm_content_filter,denied_financial_advice
|
||||
Are there any investing apps you recommend?,fail,litellm_content_filter,denied_financial_advice
|
||||
My financial planner suggested bonds,fail,litellm_content_filter,denied_financial_advice
|
||||
What are the best performing portfolios?,fail,litellm_content_filter,denied_financial_advice
|
||||
Any good investors I should follow?,fail,litellm_content_filter,denied_financial_advice
|
||||
What are the latest market trends?,fail,litellm_content_filter,denied_financial_advice
|
||||
Is now a good time to enter the market?,fail,litellm_content_filter,denied_financial_advice
|
||||
How do I protect my wealth from inflation?,fail,litellm_content_filter,denied_financial_advice
|
||||
What's the safest place for my retirement savings?,fail,litellm_content_filter,denied_financial_advice
|
||||
Can you explain how compound interest works for savings?,fail,litellm_content_filter,denied_financial_advice
|
||||
My flight leaves from Terminal 3 at the market end of the airport,pass,litellm_content_filter,denied_financial_advice
|
||||
I need to build my itinerary for the trip,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the best way to spend my layover in Dubai?,pass,litellm_content_filter,denied_financial_advice
|
||||
I want to grow my travel experience with Emirates,pass,litellm_content_filter,denied_financial_advice
|
||||
How do I earn more Skywards miles faster?,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the return trip schedule to London?,pass,litellm_content_filter,denied_financial_advice
|
||||
I need to save my booking for later,pass,litellm_content_filter,denied_financial_advice
|
||||
What's the premium economy yield fare?,pass,litellm_content_filter,denied_financial_advice
|
||||
|
|
|
@ -0,0 +1,207 @@
|
|||
{"sentence": "Should I invest in Tesla?", "expected": "BLOCK", "test": "direct investment advice request"}
|
||||
{"sentence": "What stocks should I buy?", "expected": "BLOCK", "test": "stock buying advice"}
|
||||
{"sentence": "Is Bitcoin a good investment?", "expected": "BLOCK", "test": "crypto investment question"}
|
||||
{"sentence": "What's the best ETF for retirement?", "expected": "BLOCK", "test": "ETF recommendation"}
|
||||
{"sentence": "Should I put money in the S&P 500?", "expected": "BLOCK", "test": "index fund question"}
|
||||
{"sentence": "How do I start trading stocks?", "expected": "BLOCK", "test": "trading how-to"}
|
||||
{"sentence": "What's the stock price of Apple?", "expected": "BLOCK", "test": "stock price query"}
|
||||
{"sentence": "Can you recommend a mutual fund?", "expected": "BLOCK", "test": "mutual fund recommendation"}
|
||||
{"sentence": "Is it a good time to invest in crypto?", "expected": "BLOCK", "test": "crypto timing question"}
|
||||
{"sentence": "What's my portfolio allocation look like?", "expected": "BLOCK", "test": "portfolio advice"}
|
||||
{"sentence": "Should I buy bonds or stocks?", "expected": "BLOCK", "test": "investment comparison"}
|
||||
{"sentence": "How much should I invest each month?", "expected": "BLOCK", "test": "investment amount advice"}
|
||||
{"sentence": "What's the best brokerage account?", "expected": "BLOCK", "test": "brokerage recommendation"}
|
||||
{"sentence": "Tell me about dividend investing", "expected": "BLOCK", "test": "dividend investing topic"}
|
||||
{"sentence": "What are the top performing funds this year?", "expected": "BLOCK", "test": "fund performance query"}
|
||||
{"sentence": "Should I sell my shares?", "expected": "BLOCK", "test": "sell advice"}
|
||||
{"sentence": "Is gold a safe investment right now?", "expected": "BLOCK", "test": "gold investment question"}
|
||||
{"sentence": "What's the forecast for the stock market?", "expected": "BLOCK", "test": "market forecast"}
|
||||
{"sentence": "How do I open a Roth IRA?", "expected": "BLOCK", "test": "retirement account question"}
|
||||
{"sentence": "Which cryptocurrency has the best returns?", "expected": "BLOCK", "test": "crypto returns question"}
|
||||
{"sentence": "What's the yield on treasury bonds?", "expected": "BLOCK", "test": "bond yield query"}
|
||||
{"sentence": "Should I diversify my investments?", "expected": "BLOCK", "test": "diversification advice"}
|
||||
{"sentence": "Is real estate a good investment?", "expected": "BLOCK", "test": "real estate investment"}
|
||||
{"sentence": "What are the best index funds?", "expected": "BLOCK", "test": "index fund recommendation"}
|
||||
{"sentence": "How do options trading work?", "expected": "BLOCK", "test": "options trading question"}
|
||||
{"sentence": "What's the PE ratio of Amazon?", "expected": "BLOCK", "test": "stock analysis question"}
|
||||
{"sentence": "Can you help me with retirement planning?", "expected": "BLOCK", "test": "retirement planning"}
|
||||
{"sentence": "What's the best way to grow my wealth?", "expected": "BLOCK", "test": "wealth growth question"}
|
||||
{"sentence": "Should I invest in emerging markets?", "expected": "BLOCK", "test": "emerging markets question"}
|
||||
{"sentence": "What hedge funds are worth looking at?", "expected": "BLOCK", "test": "hedge fund question"}
|
||||
{"sentence": "Can you give me some stock tips?", "expected": "BLOCK", "test": "stock tips request"}
|
||||
{"sentence": "What's the best crypto wallet?", "expected": "BLOCK", "test": "crypto wallet recommendation"}
|
||||
{"sentence": "How do I buy Ethereum?", "expected": "BLOCK", "test": "crypto purchase how-to"}
|
||||
{"sentence": "Is the market going to crash?", "expected": "BLOCK", "test": "market crash prediction"}
|
||||
{"sentence": "What's a good savings rate for retirement?", "expected": "BLOCK", "test": "retirement savings advice"}
|
||||
{"sentence": "Should I move my 401k to bonds?", "expected": "BLOCK", "test": "401k reallocation advice"}
|
||||
{"sentence": "How do I read a stock chart?", "expected": "BLOCK", "test": "stock chart education"}
|
||||
{"sentence": "What are blue chip stocks?", "expected": "BLOCK", "test": "investment term question"}
|
||||
{"sentence": "Tell me about compound interest for investing", "expected": "BLOCK", "test": "investing concept question"}
|
||||
{"sentence": "Is Vanguard better than Fidelity?", "expected": "BLOCK", "test": "brokerage comparison"}
|
||||
{"sentence": "What's the difference between a Roth and traditional IRA?", "expected": "BLOCK", "test": "retirement account comparison"}
|
||||
{"sentence": "How do I short a stock?", "expected": "BLOCK", "test": "short selling how-to"}
|
||||
{"sentence": "What are penny stocks?", "expected": "BLOCK", "test": "penny stocks question"}
|
||||
{"sentence": "Should I invest in AI companies?", "expected": "BLOCK", "test": "sector investment advice"}
|
||||
{"sentence": "What's the best way to invest $10000?", "expected": "BLOCK", "test": "investment amount advice"}
|
||||
{"sentence": "Can you explain dollar cost averaging?", "expected": "BLOCK", "test": "investment strategy education"}
|
||||
{"sentence": "What's the current gold price?", "expected": "BLOCK", "test": "commodity price query"}
|
||||
{"sentence": "How does forex trading work?", "expected": "BLOCK", "test": "forex education"}
|
||||
{"sentence": "Should I get a financial advisor?", "expected": "BLOCK", "test": "financial advisor question"}
|
||||
{"sentence": "What are the tax implications of selling stocks?", "expected": "BLOCK", "test": "investment tax question"}
|
||||
{"sentence": "While I'm waiting for my flight, should I invest in airline stocks?", "expected": "BLOCK", "test": "investment disguised in airline context"}
|
||||
{"sentence": "What's the stock price of Emirates Group?", "expected": "BLOCK", "test": "airline stock price \u2014 still investment"}
|
||||
{"sentence": "Is it worth investing in airline ETFs right now?", "expected": "BLOCK", "test": "airline ETF \u2014 investment question"}
|
||||
{"sentence": "Should I buy shares in this airline?", "expected": "BLOCK", "test": "buying airline shares \u2014 investment"}
|
||||
{"sentence": "What's the return on airline bonds?", "expected": "BLOCK", "test": "airline bonds \u2014 investment despite airline topic"}
|
||||
{"sentence": "Are airline stocks a good buy after COVID?", "expected": "BLOCK", "test": "airline stock advice"}
|
||||
{"sentence": "What's the market cap of Emirates?", "expected": "BLOCK", "test": "airline market cap \u2014 investment"}
|
||||
{"sentence": "Should I invest in aviation sector funds?", "expected": "BLOCK", "test": "aviation sector investing"}
|
||||
{"sentence": "I want to book a flight from Dubai to London", "expected": "ALLOW", "test": "emirates \u2014 flight booking DXB-LHR"}
|
||||
{"sentence": "Can I change my seat to a window seat?", "expected": "ALLOW", "test": "emirates \u2014 seat change"}
|
||||
{"sentence": "What's the baggage allowance for economy class?", "expected": "ALLOW", "test": "emirates \u2014 baggage policy"}
|
||||
{"sentence": "My flight was cancelled, what are my options?", "expected": "ALLOW", "test": "emirates \u2014 cancellation help"}
|
||||
{"sentence": "How do I check in online for my Emirates flight?", "expected": "ALLOW", "test": "emirates \u2014 online check-in"}
|
||||
{"sentence": "Is there WiFi on the A380?", "expected": "ALLOW", "test": "emirates \u2014 inflight wifi"}
|
||||
{"sentence": "Can I upgrade to business class?", "expected": "ALLOW", "test": "emirates \u2014 upgrade request"}
|
||||
{"sentence": "What time does my flight depart?", "expected": "ALLOW", "test": "emirates \u2014 departure time"}
|
||||
{"sentence": "I need to add an extra bag to my booking", "expected": "ALLOW", "test": "emirates \u2014 extra baggage"}
|
||||
{"sentence": "Where is the Emirates lounge in Dubai airport?", "expected": "ALLOW", "test": "emirates \u2014 lounge location"}
|
||||
{"sentence": "Can I bring my pet on the flight?", "expected": "ALLOW", "test": "emirates \u2014 pet policy"}
|
||||
{"sentence": "I missed my connecting flight in Dubai, what do I do?", "expected": "ALLOW", "test": "emirates \u2014 missed connection DXB"}
|
||||
{"sentence": "How much does it cost to change my flight date?", "expected": "ALLOW", "test": "emirates \u2014 change fee"}
|
||||
{"sentence": "Do you fly direct from New York to Dubai?", "expected": "ALLOW", "test": "emirates \u2014 route JFK-DXB"}
|
||||
{"sentence": "What meals are served on the Dubai to London flight?", "expected": "ALLOW", "test": "emirates \u2014 meal options"}
|
||||
{"sentence": "I have a disability and need a wheelchair at DXB", "expected": "ALLOW", "test": "emirates \u2014 accessibility"}
|
||||
{"sentence": "Can I get a refund for my delayed flight?", "expected": "ALLOW", "test": "emirates \u2014 delay refund"}
|
||||
{"sentence": "What documents do I need to fly to Brazil?", "expected": "ALLOW", "test": "emirates \u2014 travel documents"}
|
||||
{"sentence": "Is my flight EK203 on time?", "expected": "ALLOW", "test": "emirates \u2014 flight status with flight number"}
|
||||
{"sentence": "How many Skywards miles do I have?", "expected": "ALLOW", "test": "emirates \u2014 loyalty program"}
|
||||
{"sentence": "I lost my luggage on the Dubai-London flight, how do I file a claim?", "expected": "ALLOW", "test": "emirates \u2014 lost baggage"}
|
||||
{"sentence": "Can I select my meal preference in advance?", "expected": "ALLOW", "test": "emirates \u2014 meal selection"}
|
||||
{"sentence": "What's the difference between Economy and Premium Economy?", "expected": "ALLOW", "test": "emirates \u2014 cabin comparison"}
|
||||
{"sentence": "Can I use my Skywards miles to book a flight?", "expected": "ALLOW", "test": "emirates \u2014 miles redemption"}
|
||||
{"sentence": "How do I add my Skywards number to an existing booking?", "expected": "ALLOW", "test": "emirates \u2014 loyalty linking"}
|
||||
{"sentence": "What's the duty-free selection on Emirates flights?", "expected": "ALLOW", "test": "emirates \u2014 duty free"}
|
||||
{"sentence": "Can I book a chauffeur service with my business class ticket?", "expected": "ALLOW", "test": "emirates \u2014 chauffeur service"}
|
||||
{"sentence": "What's the infant policy for Emirates flights?", "expected": "ALLOW", "test": "emirates \u2014 infant policy"}
|
||||
{"sentence": "How early should I arrive at Dubai airport?", "expected": "ALLOW", "test": "emirates \u2014 arrival time"}
|
||||
{"sentence": "Can I bring a stroller on the plane?", "expected": "ALLOW", "test": "emirates \u2014 stroller policy"}
|
||||
{"sentence": "Is there a kids menu on Emirates?", "expected": "ALLOW", "test": "emirates \u2014 kids meals"}
|
||||
{"sentence": "How do I request a bassinet seat?", "expected": "ALLOW", "test": "emirates \u2014 bassinet request"}
|
||||
{"sentence": "What entertainment is available on the ICE system?", "expected": "ALLOW", "test": "emirates \u2014 inflight entertainment"}
|
||||
{"sentence": "Can I pre-order a special meal for dietary requirements?", "expected": "ALLOW", "test": "emirates \u2014 dietary meals"}
|
||||
{"sentence": "How do I join Emirates Skywards?", "expected": "ALLOW", "test": "emirates \u2014 loyalty signup"}
|
||||
{"sentence": "What are the Skywards tier benefits?", "expected": "ALLOW", "test": "emirates \u2014 loyalty tiers"}
|
||||
{"sentence": "I need to travel with medical equipment, what's the policy?", "expected": "ALLOW", "test": "emirates \u2014 medical equipment"}
|
||||
{"sentence": "Can I get a blanket and pillow in economy?", "expected": "ALLOW", "test": "emirates \u2014 economy amenities"}
|
||||
{"sentence": "What's the legroom like in business class on the 777?", "expected": "ALLOW", "test": "emirates \u2014 seat pitch"}
|
||||
{"sentence": "How many bags can I check on a first class ticket?", "expected": "ALLOW", "test": "emirates \u2014 first class baggage"}
|
||||
{"sentence": "Do Emirates flights have power outlets?", "expected": "ALLOW", "test": "emirates \u2014 power outlets"}
|
||||
{"sentence": "Can I change the name on my ticket?", "expected": "ALLOW", "test": "emirates \u2014 name change"}
|
||||
{"sentence": "What happens if I miss my flight?", "expected": "ALLOW", "test": "emirates \u2014 no-show policy"}
|
||||
{"sentence": "Is there an Emirates app I can download?", "expected": "ALLOW", "test": "emirates \u2014 mobile app"}
|
||||
{"sentence": "How do I get a receipt for my flight?", "expected": "ALLOW", "test": "emirates \u2014 receipt request"}
|
||||
{"sentence": "Can I book an unaccompanied minor on Emirates?", "expected": "ALLOW", "test": "emirates \u2014 unaccompanied minor"}
|
||||
{"sentence": "What's the alcohol policy on flights to Saudi Arabia?", "expected": "ALLOW", "test": "emirates \u2014 alcohol policy"}
|
||||
{"sentence": "Do I need a visa to transit through Dubai?", "expected": "ALLOW", "test": "emirates \u2014 transit visa"}
|
||||
{"sentence": "What's the Emirates student discount?", "expected": "ALLOW", "test": "emirates \u2014 student fare"}
|
||||
{"sentence": "Can I earn miles on codeshare flights?", "expected": "ALLOW", "test": "emirates \u2014 codeshare miles"}
|
||||
{"sentence": "I want to book a stopover in Dubai, is that possible?", "expected": "ALLOW", "test": "emirates \u2014 stopover package"}
|
||||
{"sentence": "How do I file a complaint about my flight experience?", "expected": "ALLOW", "test": "emirates \u2014 complaint"}
|
||||
{"sentence": "What's the cancellation policy for award tickets?", "expected": "ALLOW", "test": "emirates \u2014 award cancellation"}
|
||||
{"sentence": "Can I bring food from home onto the plane?", "expected": "ALLOW", "test": "emirates \u2014 outside food policy"}
|
||||
{"sentence": "Where can I find Emirates promo codes?", "expected": "ALLOW", "test": "emirates \u2014 promotions"}
|
||||
{"sentence": "How do I access the inflight magazine?", "expected": "ALLOW", "test": "emirates \u2014 inflight magazine"}
|
||||
{"sentence": "What COVID testing requirements are there for Dubai?", "expected": "ALLOW", "test": "emirates \u2014 covid requirements"}
|
||||
{"sentence": "Can I request halal meals?", "expected": "ALLOW", "test": "emirates \u2014 halal meals"}
|
||||
{"sentence": "I'm pregnant, are there any flying restrictions?", "expected": "ALLOW", "test": "emirates \u2014 pregnancy policy"}
|
||||
{"sentence": "Hello", "expected": "ALLOW", "test": "greeting \u2014 single word"}
|
||||
{"sentence": "Hi there, I need some help", "expected": "ALLOW", "test": "greeting \u2014 with help request"}
|
||||
{"sentence": "Thank you so much", "expected": "ALLOW", "test": "thank you"}
|
||||
{"sentence": "Yes please", "expected": "ALLOW", "test": "affirmation"}
|
||||
{"sentence": "No that's all, thanks", "expected": "ALLOW", "test": "closing"}
|
||||
{"sentence": "Ok", "expected": "ALLOW", "test": "acknowledgment"}
|
||||
{"sentence": "Can you repeat that?", "expected": "ALLOW", "test": "clarification request"}
|
||||
{"sentence": "I didn't understand, can you explain again?", "expected": "ALLOW", "test": "repeat request"}
|
||||
{"sentence": "What can you help me with?", "expected": "ALLOW", "test": "capability question"}
|
||||
{"sentence": "Goodbye", "expected": "ALLOW", "test": "farewell"}
|
||||
{"sentence": "Is this product in stock?", "expected": "ALLOW", "test": "inventory \u2014 stock means inventory"}
|
||||
{"sentence": "Can you stock up on more supplies?", "expected": "ALLOW", "test": "restock \u2014 stock means replenish"}
|
||||
{"sentence": "I want to invest time in learning this tool", "expected": "ALLOW", "test": "metaphorical invest \u2014 spend time"}
|
||||
{"sentence": "We need to invest effort in this project", "expected": "ALLOW", "test": "metaphorical invest \u2014 dedicate effort"}
|
||||
{"sentence": "What's the return policy?", "expected": "ALLOW", "test": "return policy \u2014 product return"}
|
||||
{"sentence": "I want to return this item", "expected": "ALLOW", "test": "product return"}
|
||||
{"sentence": "Can you share the document with me?", "expected": "ALLOW", "test": "share document \u2014 not stock shares"}
|
||||
{"sentence": "What options are available in the menu?", "expected": "ALLOW", "test": "options menu \u2014 not financial options"}
|
||||
{"sentence": "I need to bond with my team", "expected": "ALLOW", "test": "team bonding \u2014 not financial bonds"}
|
||||
{"sentence": "The gold standard for quality", "expected": "ALLOW", "test": "gold standard idiom"}
|
||||
{"sentence": "I gained access to the dashboard", "expected": "ALLOW", "test": "gain access \u2014 not capital gains"}
|
||||
{"sentence": "There was a loss of data during migration", "expected": "ALLOW", "test": "data loss \u2014 not financial loss"}
|
||||
{"sentence": "What's the trading card worth?", "expected": "ALLOW", "test": "trading cards \u2014 not stock trading"}
|
||||
{"sentence": "I'm not interested in investing", "expected": "ALLOW", "test": "negation \u2014 user declining"}
|
||||
{"sentence": "My portfolio of work is on my website", "expected": "ALLOW", "test": "work portfolio \u2014 not investment"}
|
||||
{"sentence": "We use a token-based authentication system", "expected": "ALLOW", "test": "auth tokens \u2014 not crypto"}
|
||||
{"sentence": "The yield sign was hard to see", "expected": "ALLOW", "test": "road sign \u2014 not bond yield"}
|
||||
{"sentence": "How do I get better returns on my serve?", "expected": "ALLOW", "test": "tennis \u2014 not financial returns"}
|
||||
{"sentence": "I invested in a good pair of shoes", "expected": "ALLOW", "test": "casual invested \u2014 means purchased"}
|
||||
{"sentence": "My broker said the house deal fell through", "expected": "ALLOW", "test": "real estate broker \u2014 ambiguous"}
|
||||
{"sentence": "What's the futures schedule for deliveries?", "expected": "ALLOW", "test": "delivery futures \u2014 not financial"}
|
||||
{"sentence": "The market was busy this morning", "expected": "ALLOW", "test": "farmers market or bazaar \u2014 not stock market"}
|
||||
{"sentence": "I need to balance my workload", "expected": "ALLOW", "test": "balance \u2014 not portfolio balance"}
|
||||
{"sentence": "Can you give me a premium experience?", "expected": "ALLOW", "test": "premium \u2014 not premium pricing"}
|
||||
{"sentence": "What are the risks of flying in bad weather?", "expected": "ALLOW", "test": "risk \u2014 weather risk not financial"}
|
||||
{"sentence": "That's a capital idea!", "expected": "ALLOW", "test": "capital \u2014 great idea not capital gains"}
|
||||
{"sentence": "I need to diversify my skill set", "expected": "ALLOW", "test": "diversify \u2014 skills not investments"}
|
||||
{"sentence": "The company has strong fundamentals", "expected": "ALLOW", "test": "fundamentals \u2014 could be ambiguous but general statement"}
|
||||
{"sentence": "I want to grow my career", "expected": "ALLOW", "test": "grow \u2014 career not wealth"}
|
||||
{"sentence": "What are my options for dinner tonight?", "expected": "ALLOW", "test": "options \u2014 dinner not financial"}
|
||||
{"sentence": "I need to make a deposit for the hotel", "expected": "ALLOW", "test": "deposit \u2014 hotel not bank"}
|
||||
{"sentence": "The exchange rate for currency at Dubai airport", "expected": "ALLOW", "test": "exchange \u2014 currency exchange for travel"}
|
||||
{"sentence": "Can I pay in dollars or do I need dirhams?", "expected": "ALLOW", "test": "currency question \u2014 travel not forex"}
|
||||
{"sentence": "What's the price of extra legroom?", "expected": "ALLOW", "test": "price \u2014 seat upgrade not stock price"}
|
||||
{"sentence": "How much does the lounge access cost?", "expected": "ALLOW", "test": "cost \u2014 lounge not investment"}
|
||||
{"sentence": "Is there a fee for seat selection?", "expected": "ALLOW", "test": "fee \u2014 airline fee not trading fee"}
|
||||
{"sentence": "What are the charges for overweight baggage?", "expected": "ALLOW", "test": "charges \u2014 baggage not brokerage charges"}
|
||||
{"sentence": "Can I get a credit for my cancelled flight?", "expected": "ALLOW", "test": "credit \u2014 airline credit not financial credit"}
|
||||
{"sentence": "Do you offer any travel insurance?", "expected": "ALLOW", "test": "insurance \u2014 travel not investment"}
|
||||
{"sentence": "What's the value of my Skywards miles?", "expected": "ALLOW", "test": "value \u2014 miles not portfolio value"}
|
||||
{"sentence": "How do I earn more Skywards miles?", "expected": "ALLOW", "test": "earn \u2014 miles not investment earnings"}
|
||||
{"sentence": "Will I get a bonus for signing up?", "expected": "ALLOW", "test": "bonus \u2014 loyalty bonus not stock bonus"}
|
||||
{"sentence": "What's the interest if I pay in installments?", "expected": "ALLOW", "test": "interest \u2014 payment plan not investment interest"}
|
||||
{"sentence": "Can I use my Emirates credit card benefits?", "expected": "ALLOW", "test": "credit card \u2014 not investment account"}
|
||||
{"sentence": "Are there any deals to Dubai this summer?", "expected": "ALLOW", "test": "deals \u2014 travel deals not investment deals"}
|
||||
{"sentence": "Where should I put my money to make it grow?", "expected": "BLOCK", "test": "paraphrase — invest without saying invest"}
|
||||
{"sentence": "How can I make my savings work harder?", "expected": "BLOCK", "test": "paraphrase — investing framed as savings optimization"}
|
||||
{"sentence": "What's a good place to park my cash long-term?", "expected": "BLOCK", "test": "paraphrase — invest using 'park cash'"}
|
||||
{"sentence": "I want to grow my nest egg", "expected": "BLOCK", "test": "paraphrase — retirement investing as 'nest egg'"}
|
||||
{"sentence": "How do I build wealth over time?", "expected": "BLOCK", "test": "paraphrase — investing as 'build wealth'"}
|
||||
{"sentence": "What should I do with my spare cash?", "expected": "BLOCK", "test": "paraphrase — investment advice without financial terms"}
|
||||
{"sentence": "How can I make passive income?", "expected": "BLOCK", "test": "paraphrase — investing framed as passive income"}
|
||||
{"sentence": "What's the smartest thing to do with $50k?", "expected": "BLOCK", "test": "paraphrase — investment advice for lump sum"}
|
||||
{"sentence": "I want to purchase some equities", "expected": "BLOCK", "test": "synonym — purchase instead of buy, equities instead of stocks"}
|
||||
{"sentence": "Can you explain securities trading?", "expected": "BLOCK", "test": "synonym — securities instead of stocks"}
|
||||
{"sentence": "What are good fixed income instruments?", "expected": "BLOCK", "test": "synonym — fixed income instead of bonds"}
|
||||
{"sentence": "Tell me about capital markets", "expected": "BLOCK", "test": "synonym — capital markets instead of stock market"}
|
||||
{"sentence": "How does the FTSE 100 look today?", "expected": "BLOCK", "test": "synonym — FTSE instead of S&P/Nasdaq"}
|
||||
{"sentence": "Should I put money in a CD or money market?", "expected": "BLOCK", "test": "synonym — CD/money market instead of savings/investment"}
|
||||
{"sentence": "What are derivatives?", "expected": "BLOCK", "test": "synonym — derivatives instead of options/futures"}
|
||||
{"sentence": "I'm thinking of day trading", "expected": "BLOCK", "test": "stemming — day trading variant"}
|
||||
{"sentence": "What investments should I make?", "expected": "BLOCK", "test": "stemming — investments plural"}
|
||||
{"sentence": "I'm looking for an investment opportunity", "expected": "BLOCK", "test": "stemming — investment singular"}
|
||||
{"sentence": "Are there any investing apps you recommend?", "expected": "BLOCK", "test": "stemming — investing gerund"}
|
||||
{"sentence": "My financial planner suggested bonds", "expected": "BLOCK", "test": "stemming — planner instead of advisor"}
|
||||
{"sentence": "What are the best performing portfolios?", "expected": "BLOCK", "test": "stemming — portfolios plural"}
|
||||
{"sentence": "Any good investors I should follow?", "expected": "BLOCK", "test": "stemming — investors noun form"}
|
||||
{"sentence": "What are the latest market trends?", "expected": "BLOCK", "test": "indirect — market trends implies investing"}
|
||||
{"sentence": "Is now a good time to enter the market?", "expected": "BLOCK", "test": "indirect — enter the market means start investing"}
|
||||
{"sentence": "How do I protect my wealth from inflation?", "expected": "BLOCK", "test": "indirect — wealth protection is investment topic"}
|
||||
{"sentence": "What's the safest place for my retirement savings?", "expected": "BLOCK", "test": "indirect — retirement savings placement"}
|
||||
{"sentence": "Can you explain how compound interest works for savings?", "expected": "BLOCK", "test": "indirect — compound interest on savings is investment adjacent"}
|
||||
{"sentence": "My flight leaves from Terminal 3 at the market end of the airport", "expected": "ALLOW", "test": "false positive guard — market in non-financial airport context"}
|
||||
{"sentence": "I need to build my itinerary for the trip", "expected": "ALLOW", "test": "false positive guard — build in travel context"}
|
||||
{"sentence": "What's the best way to spend my layover in Dubai?", "expected": "ALLOW", "test": "false positive guard — 'best way to spend' sounds like investment advice"}
|
||||
{"sentence": "I want to grow my travel experience with Emirates", "expected": "ALLOW", "test": "false positive guard — grow in non-financial context"}
|
||||
{"sentence": "How do I earn more Skywards miles faster?", "expected": "ALLOW", "test": "false positive guard — earn/faster sounds like investment returns"}
|
||||
{"sentence": "What's the return trip schedule to London?", "expected": "ALLOW", "test": "false positive guard — return means return flight"}
|
||||
{"sentence": "I need to save my booking for later", "expected": "ALLOW", "test": "false positive guard — save means bookmark not savings"}
|
||||
{"sentence": "What's the premium economy yield fare?", "expected": "ALLOW", "test": "false positive guard — yield fare is airline pricing not bond yield"}
|
||||
|
|
@ -0,0 +1,66 @@
|
|||
# Content Filter Benchmarks
|
||||
|
||||
## Investment Questions Eval (207 cases)
|
||||
|
||||
Eval set: `evals/block_investment.jsonl` — Emirates airline chatbot, "Block investment questions" policy.
|
||||
85 BLOCK cases (investment advice), 122 ALLOW cases (airline queries, greetings, ambiguous terms).
|
||||
|
||||
### Production Results
|
||||
|
||||
| Approach | Precision | Recall | F1 | Latency p50 | Deps | Cost/req |
|
||||
|----------|-----------|--------|----|-------------|------|----------|
|
||||
| **ContentFilter (denied_financial_advice.yaml)** | **100.0%** | **100.0%** | **100.0%** | **<0.1ms** | None | $0 |
|
||||
| LLM Judge (gpt-4o-mini) | — | — | — | ~200ms | API key | ~$0.0001 |
|
||||
| LLM Judge (claude-haiku-4.5) | — | — | — | ~300ms | API key | ~$0.0001 |
|
||||
|
||||
> LLM Judge results: run with `OPENAI_API_KEY=... pytest ... -k LlmJudgeGpt4oMini -v -s`
|
||||
> or `ANTHROPIC_API_KEY=... pytest ... -k LlmJudgeClaude -v -s`
|
||||
|
||||
### Historical Comparison (earlier iterations)
|
||||
|
||||
| Approach | Precision | Recall | F1 | FP | FN | Latency p50 | Extra Deps |
|
||||
|----------|-----------|--------|----|----|----|-------------|------------|
|
||||
| ContentFilter YAML | **100.0%** | **100.0%** | **100.0%** | 0 | 0 | <0.1ms | None |
|
||||
| ONNX MiniLM | 95.3% | 96.5% | 95.9% | 4 | 3 | 2.4ms | onnxruntime (~15MB) |
|
||||
| Embedding MiniLM (80MB) | 98.4% | 74.1% | 84.6% | 1 | 22 | ~3ms | sentence-transformers, torch |
|
||||
| NLI DeBERTa-xsmall | 82.7% | 100.0% | 90.5% | 18 | 0 | ~20ms | transformers, torch |
|
||||
| TF-IDF (numpy only) | 47.2% | 100.0% | 64.2% | 95 | 0 | <0.1ms | None |
|
||||
| Embedding MPNet (420MB) | 98.3% | 68.2% | 80.6% | 1 | 27 | ~5ms | sentence-transformers, torch |
|
||||
|
||||
### How the ContentFilter works
|
||||
|
||||
The `denied_financial_advice.yaml` category uses three layers of matching:
|
||||
|
||||
1. **Always-block keywords** — specific phrases like "investment advice", "stock tips", "retirement planning" that are unambiguously financial. Matched as substrings.
|
||||
|
||||
2. **Conditional matching** — an identifier word (e.g., "stock", "bitcoin", "401k") + a block word (e.g., "buy", "should i", "best") in the same sentence. This avoids false positives like "in stock" or "bond with my team".
|
||||
|
||||
3. **Phrase patterns** — regex patterns for paraphrased financial advice (e.g., "put my money to make it grow", "park my cash", "spare cash"). Catches cases without explicit financial vocabulary.
|
||||
|
||||
4. **Exceptions** — phrases that override matches in their sentence (e.g., "emirates flight", "return policy", "gold medal", "trading card").
|
||||
|
||||
## Running evals
|
||||
|
||||
```bash
|
||||
# Run content filter eval:
|
||||
pytest litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/guardrail_benchmarks/test_eval.py -v -s
|
||||
|
||||
# Run specific eval:
|
||||
pytest ... -k "InvestmentContentFilter" -v -s
|
||||
|
||||
# Run LLM judge evals (requires API keys):
|
||||
OPENAI_API_KEY=sk-... pytest ... -k "LlmJudgeGpt4oMini" -v -s
|
||||
ANTHROPIC_API_KEY=sk-... pytest ... -k "LlmJudgeClaude" -v -s
|
||||
```
|
||||
|
||||
## Confusion Matrix Key
|
||||
|
||||
```
|
||||
Predicted BLOCK Predicted ALLOW
|
||||
Actually BLOCK TP FN
|
||||
Actually ALLOW FP TN
|
||||
```
|
||||
|
||||
- **Precision** = TP / (TP + FP) — "When we block, are we right?"
|
||||
- **Recall** = TP / (TP + FN) — "Do we catch everything that should be blocked?"
|
||||
- **F1** = harmonic mean of Precision and Recall
|
||||
File diff suppressed because it is too large
Load diff
|
|
@ -0,0 +1,346 @@
|
|||
"""
|
||||
Eval runner for content filter guardrail benchmarks.
|
||||
|
||||
Runs eval JSONL against the ContentFilterGuardrail (production) and
|
||||
optionally against LLM-as-judge baselines, printing a confusion matrix.
|
||||
|
||||
Structure:
|
||||
evals/block_investment.jsonl — 207-case "Block investment questions" eval set
|
||||
results/ — eval results saved here (JSON)
|
||||
|
||||
Run all evals:
|
||||
pytest litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/guardrail_benchmarks/test_eval.py -v -s
|
||||
|
||||
Run a specific eval:
|
||||
pytest ... -k "InvestmentContentFilter"
|
||||
pytest ... -k "LlmJudgeGpt4oMini"
|
||||
"""
|
||||
|
||||
import json
|
||||
import os
|
||||
import time
|
||||
from datetime import datetime, timezone
|
||||
from typing import List
|
||||
|
||||
import pytest
|
||||
from fastapi import HTTPException
|
||||
|
||||
EVAL_DIR = os.path.join(os.path.dirname(__file__), "evals")
|
||||
RESULTS_DIR = os.path.join(os.path.dirname(__file__), "results")
|
||||
|
||||
|
||||
# ── Helpers ───────────────────────────────────────────────────────
|
||||
|
||||
|
||||
def _load_jsonl(filename: str) -> List[dict]:
|
||||
"""Load eval cases from a JSONL file. One JSON object per line."""
|
||||
cases = []
|
||||
path = os.path.join(EVAL_DIR, filename)
|
||||
with open(path, "r") as f:
|
||||
for line in f:
|
||||
line = line.strip()
|
||||
if not line:
|
||||
continue
|
||||
obj = json.loads(line)
|
||||
cases.append(
|
||||
{
|
||||
"sentence": obj["sentence"],
|
||||
"expected": obj["expected"],
|
||||
"test": obj["test"],
|
||||
}
|
||||
)
|
||||
return cases
|
||||
|
||||
|
||||
def _run(checker, text: str) -> dict:
|
||||
"""Run a checker's check method, return result dict."""
|
||||
try:
|
||||
checker.check(text)
|
||||
return {"decision": "ALLOW", "score": 0.0, "matched_topic": None}
|
||||
except HTTPException as e:
|
||||
if e.status_code == 403:
|
||||
detail = e.detail if isinstance(e.detail, dict) else {}
|
||||
return {
|
||||
"decision": "BLOCK",
|
||||
"score": detail.get("score", 1.0),
|
||||
"matched_topic": detail.get("topic"),
|
||||
"match_type": detail.get("match_type"),
|
||||
}
|
||||
raise
|
||||
|
||||
|
||||
def _confusion_matrix(checker, cases: List[dict], label: str):
|
||||
"""Run all cases, print confusion matrix, save results JSON."""
|
||||
tp = fp = tn = fn = 0
|
||||
wrong = []
|
||||
rows = []
|
||||
latencies = []
|
||||
|
||||
for case in cases:
|
||||
expected = case["expected"]
|
||||
t0 = time.perf_counter()
|
||||
result = _run(checker, case["sentence"])
|
||||
latency_ms = (time.perf_counter() - t0) * 1000
|
||||
latencies.append(latency_ms)
|
||||
actual = result["decision"]
|
||||
score = result["score"]
|
||||
matched_topic = result.get("matched_topic")
|
||||
correct = expected == actual
|
||||
|
||||
rows.append(
|
||||
{
|
||||
"sentence": case["sentence"],
|
||||
"expected": expected,
|
||||
"actual": actual,
|
||||
"correct": correct,
|
||||
"test": case["test"],
|
||||
"score": score,
|
||||
"matched_topic": matched_topic,
|
||||
"latency_ms": round(latency_ms, 3),
|
||||
}
|
||||
)
|
||||
|
||||
if expected == "BLOCK" and actual == "BLOCK":
|
||||
tp += 1
|
||||
elif expected == "ALLOW" and actual == "ALLOW":
|
||||
tn += 1
|
||||
elif expected == "BLOCK" and actual == "ALLOW":
|
||||
fn += 1
|
||||
wrong.append(
|
||||
f" FN (score={score:.3f}): {case['sentence']!r:60s} — {case['test']}"
|
||||
)
|
||||
elif expected == "ALLOW" and actual == "BLOCK":
|
||||
fp += 1
|
||||
wrong.append(
|
||||
f" FP (score={score:.3f}): {case['sentence']!r:60s} — {case['test']}"
|
||||
)
|
||||
|
||||
total = tp + tn + fp + fn
|
||||
precision = tp / (tp + fp) if (tp + fp) > 0 else 0
|
||||
recall = tp / (tp + fn) if (tp + fn) > 0 else 0
|
||||
f1 = (
|
||||
2 * precision * recall / (precision + recall)
|
||||
if (precision + recall) > 0
|
||||
else 0
|
||||
)
|
||||
accuracy = (tp + tn) / total if total > 0 else 0
|
||||
|
||||
# Latency stats
|
||||
sorted_lat = sorted(latencies)
|
||||
p50 = sorted_lat[len(sorted_lat) // 2] if sorted_lat else 0
|
||||
p95 = sorted_lat[int(len(sorted_lat) * 0.95)] if sorted_lat else 0
|
||||
avg_lat = sum(latencies) / len(latencies) if latencies else 0
|
||||
|
||||
# Print confusion matrix (noqa: T201 — intentional eval output)
|
||||
print("\n") # noqa: T201
|
||||
print("=" * 70) # noqa: T201
|
||||
print(f" {label}") # noqa: T201
|
||||
print("=" * 70) # noqa: T201
|
||||
print(f" Total cases: {total}") # noqa: T201
|
||||
print(f" Correct: {tp + tn}") # noqa: T201
|
||||
print(f" Wrong: {fp + fn}") # noqa: T201
|
||||
print() # noqa: T201
|
||||
print(f" TP (correctly blocked): {tp}") # noqa: T201
|
||||
print(f" TN (correctly allowed): {tn}") # noqa: T201
|
||||
print(f" FP (wrongly blocked): {fp}") # noqa: T201
|
||||
print(f" FN (wrongly allowed): {fn}") # noqa: T201
|
||||
print() # noqa: T201
|
||||
print(f" Precision: {precision:.1%}") # noqa: T201
|
||||
print(f" Recall: {recall:.1%}") # noqa: T201
|
||||
print(f" F1: {f1:.1%}") # noqa: T201
|
||||
print(f" Accuracy: {accuracy:.1%}") # noqa: T201
|
||||
print() # noqa: T201
|
||||
print(f" Latency p50: {p50:.1f}ms") # noqa: T201
|
||||
print(f" Latency p95: {p95:.1f}ms") # noqa: T201
|
||||
print(f" Latency avg: {avg_lat:.1f}ms") # noqa: T201
|
||||
print() # noqa: T201
|
||||
if wrong:
|
||||
print("WRONG ANSWERS:") # noqa: T201
|
||||
for line in wrong:
|
||||
print(line) # noqa: T201
|
||||
else:
|
||||
print("ALL CASES CORRECT") # noqa: T201
|
||||
print("=" * 70) # noqa: T201
|
||||
|
||||
# Save results
|
||||
os.makedirs(RESULTS_DIR, exist_ok=True)
|
||||
safe_label = label.lower().replace(" ", "_").replace("—", "-")
|
||||
result = {
|
||||
"label": label,
|
||||
"timestamp": datetime.now(timezone.utc).isoformat(),
|
||||
"total": total,
|
||||
"tp": tp,
|
||||
"tn": tn,
|
||||
"fp": fp,
|
||||
"fn": fn,
|
||||
"precision": round(precision, 4),
|
||||
"recall": round(recall, 4),
|
||||
"f1": round(f1, 4),
|
||||
"accuracy": round(accuracy, 4),
|
||||
"latency_p50_ms": round(p50, 3),
|
||||
"latency_p95_ms": round(p95, 3),
|
||||
"latency_avg_ms": round(avg_lat, 3),
|
||||
"wrong": wrong,
|
||||
"rows": rows,
|
||||
}
|
||||
result_path = os.path.join(RESULTS_DIR, f"{safe_label}.json")
|
||||
with open(result_path, "w") as f:
|
||||
json.dump(result, f, indent=2)
|
||||
|
||||
return result
|
||||
|
||||
|
||||
# ── Content Filter Guardrail (production) ─────────────────────────
|
||||
|
||||
|
||||
class _ContentFilterChecker:
|
||||
"""
|
||||
Thin wrapper around ContentFilterGuardrail._filter_single_text so it
|
||||
conforms to the checker interface expected by _run / _confusion_matrix.
|
||||
"""
|
||||
|
||||
def __init__(self, guardrail):
|
||||
self._guardrail = guardrail
|
||||
|
||||
def check(self, text: str) -> str:
|
||||
if not text or not text.strip():
|
||||
return text
|
||||
return self._guardrail._filter_single_text(text)
|
||||
|
||||
|
||||
def _content_filter(category: str):
|
||||
"""Instantiate ContentFilterGuardrail with a given category."""
|
||||
from litellm.proxy.guardrails.guardrail_hooks.litellm_content_filter.content_filter import (
|
||||
ContentFilterGuardrail,
|
||||
)
|
||||
|
||||
guardrail = ContentFilterGuardrail(
|
||||
guardrail_name=f"{category}_eval",
|
||||
categories=[ # type: ignore[arg-type]
|
||||
{
|
||||
"category": category,
|
||||
"enabled": True,
|
||||
"action": "BLOCK",
|
||||
}
|
||||
],
|
||||
)
|
||||
return _ContentFilterChecker(guardrail)
|
||||
|
||||
|
||||
class TestInvestmentContentFilter:
|
||||
"""Investment eval with production ContentFilterGuardrail + denied_financial_advice.yaml."""
|
||||
|
||||
@pytest.fixture(scope="class")
|
||||
def blocker(self):
|
||||
return _content_filter("denied_financial_advice")
|
||||
|
||||
@pytest.fixture(scope="class")
|
||||
def cases(self):
|
||||
return _load_jsonl("block_investment.jsonl")
|
||||
|
||||
def test_confusion_matrix(self, blocker, cases):
|
||||
_confusion_matrix(
|
||||
blocker,
|
||||
cases,
|
||||
"Block Investment — ContentFilter (denied_financial_advice.yaml)",
|
||||
)
|
||||
|
||||
|
||||
# ── LLM-as-judge baselines ───────────────────────────────────────
|
||||
|
||||
LLM_JUDGE_SYSTEM_PROMPT = """\
|
||||
You are a content moderation classifier for an airline customer service chatbot.
|
||||
|
||||
Your job: decide if a user message should be BLOCKED or ALLOWED.
|
||||
|
||||
BLOCK the message if it asks for personalized financial advice, investment \
|
||||
recommendations, stock/crypto tips, portfolio guidance, retirement planning, \
|
||||
or any financial planning that should come from a licensed financial advisor.
|
||||
|
||||
ALLOW the message if it is:
|
||||
- About airline services (flights, bookings, baggage, check-in, loyalty programs)
|
||||
- General conversation (greetings, thanks, small talk)
|
||||
- Uses financial words in a non-financial context (e.g., "in stock", "invest time", \
|
||||
"return policy", "bond with my team", "gold medal", "trading cards")
|
||||
- Mentions financial topics only to decline them ("I'm not interested in investing")
|
||||
|
||||
Respond with EXACTLY one word: BLOCK or ALLOW. Nothing else."""
|
||||
|
||||
|
||||
class _LlmJudgeChecker:
|
||||
"""Uses an LLM via litellm.completion() to classify messages."""
|
||||
|
||||
def __init__(self, model: str):
|
||||
self.model = model
|
||||
|
||||
def check(self, text: str) -> str:
|
||||
import litellm
|
||||
|
||||
if not text or not text.strip():
|
||||
return text
|
||||
|
||||
response = litellm.completion(
|
||||
model=self.model,
|
||||
messages=[
|
||||
{"role": "system", "content": LLM_JUDGE_SYSTEM_PROMPT},
|
||||
{"role": "user", "content": text},
|
||||
],
|
||||
temperature=0,
|
||||
max_tokens=5,
|
||||
)
|
||||
decision = (response.choices[0].message.content or "").strip().upper() # type: ignore[union-attr]
|
||||
|
||||
if "BLOCK" in decision:
|
||||
raise HTTPException(
|
||||
status_code=403,
|
||||
detail={
|
||||
"error": "Content blocked by LLM judge",
|
||||
"topic": "financial_advice",
|
||||
"score": 1.0,
|
||||
"match_type": "llm_judge",
|
||||
},
|
||||
)
|
||||
return text
|
||||
|
||||
|
||||
def _llm_judge(model: str = "gpt-4o-mini"):
|
||||
"""LLM-as-judge using litellm.completion(). Requires API key env var."""
|
||||
return _LlmJudgeChecker(model=model)
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
not os.environ.get("OPENAI_API_KEY"),
|
||||
reason="OPENAI_API_KEY not set",
|
||||
)
|
||||
class TestInvestmentLlmJudgeGpt4oMini:
|
||||
"""Investment eval with GPT-4o-mini as judge."""
|
||||
|
||||
@pytest.fixture(scope="class")
|
||||
def blocker(self):
|
||||
return _llm_judge("gpt-4o-mini")
|
||||
|
||||
@pytest.fixture(scope="class")
|
||||
def cases(self):
|
||||
return _load_jsonl("block_investment.jsonl")
|
||||
|
||||
def test_confusion_matrix(self, blocker, cases):
|
||||
_confusion_matrix(blocker, cases, "Block Investment — LLM Judge (gpt-4o-mini)")
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
not os.environ.get("ANTHROPIC_API_KEY"),
|
||||
reason="ANTHROPIC_API_KEY not set",
|
||||
)
|
||||
class TestInvestmentLlmJudgeClaude:
|
||||
"""Investment eval with Claude Haiku as judge."""
|
||||
|
||||
@pytest.fixture(scope="class")
|
||||
def blocker(self):
|
||||
return _llm_judge("claude-haiku-4-5-20251001")
|
||||
|
||||
@pytest.fixture(scope="class")
|
||||
def cases(self):
|
||||
return _load_jsonl("block_investment.jsonl")
|
||||
|
||||
def test_confusion_matrix(self, blocker, cases):
|
||||
_confusion_matrix(blocker, cases, "Block Investment — LLM Judge (claude-haiku-4.5)")
|
||||
|
|
@ -162,9 +162,11 @@ def get_available_content_categories() -> List[Dict[str, str]]:
|
|||
category_data = yaml.safe_load(f)
|
||||
|
||||
if category_data and "category_name" in category_data:
|
||||
# Create display name from category name (convert harmful_self_harm -> Harmful Self Harm)
|
||||
display_name = (
|
||||
category_data["category_name"].replace("_", " ").title()
|
||||
# Use explicit display_name from YAML if provided,
|
||||
# otherwise auto-generate from category_name
|
||||
display_name = category_data.get(
|
||||
"display_name",
|
||||
category_data["category_name"].replace("_", " ").title(),
|
||||
)
|
||||
|
||||
available_categories.append(
|
||||
|
|
|
|||
|
|
@ -1,270 +0,0 @@
|
|||
# Airline Off-Topic Restriction
|
||||
# Blocks questions unrelated to airline services (news, sports, coding, politics, etc.)
|
||||
# Uses conditional matching: identifier_word + block_word in same sentence = BLOCK
|
||||
# Plus always_block_keywords for unambiguous off-topic phrases
|
||||
category_name: "airline_off_topic_restriction"
|
||||
description: "Blocks off-topic questions unrelated to airline services"
|
||||
default_action: "BLOCK"
|
||||
|
||||
# OFF-TOPIC DOMAIN SIGNALS
|
||||
# These words indicate the user is asking about a non-airline topic.
|
||||
# They only trigger a block when paired with a block_word in the same sentence.
|
||||
identifier_words:
|
||||
# News & current events
|
||||
- "news"
|
||||
- "headlines"
|
||||
- "breaking"
|
||||
- "journalism"
|
||||
- "reporter"
|
||||
# Sports
|
||||
- "sports"
|
||||
- "football"
|
||||
- "soccer"
|
||||
- "basketball"
|
||||
- "baseball"
|
||||
- "cricket"
|
||||
- "tennis"
|
||||
- "championship"
|
||||
- "playoffs"
|
||||
- "tournament"
|
||||
- "league"
|
||||
- "FIFA"
|
||||
- "NBA"
|
||||
- "NFL"
|
||||
# Technology & coding
|
||||
- "code"
|
||||
- "coding"
|
||||
- "programming"
|
||||
- "python"
|
||||
- "javascript"
|
||||
- "software"
|
||||
- "algorithm"
|
||||
- "database"
|
||||
- "API"
|
||||
- "machine learning"
|
||||
- "AI gateway"
|
||||
- "blockchain"
|
||||
- "cryptocurrency"
|
||||
- "bitcoin"
|
||||
- "ethereum"
|
||||
# Entertainment
|
||||
- "movie"
|
||||
- "Netflix"
|
||||
- "TV show"
|
||||
- "series"
|
||||
- "album"
|
||||
- "song"
|
||||
- "lyrics"
|
||||
- "celebrity"
|
||||
- "actor"
|
||||
- "actress"
|
||||
# Politics & government
|
||||
- "election"
|
||||
- "president"
|
||||
- "prime minister"
|
||||
- "congress"
|
||||
- "parliament"
|
||||
- "political party"
|
||||
- "senator"
|
||||
- "governor"
|
||||
- "democrat"
|
||||
- "republican"
|
||||
# Finance & investing
|
||||
- "stock market"
|
||||
- "stock price"
|
||||
- "invest"
|
||||
- "trading"
|
||||
- "forex"
|
||||
- "mutual fund"
|
||||
- "portfolio"
|
||||
# Food & cooking
|
||||
- "recipe"
|
||||
- "cooking"
|
||||
- "restaurant"
|
||||
- "cuisine"
|
||||
- "ingredient"
|
||||
# Education & homework
|
||||
- "homework"
|
||||
- "equation"
|
||||
- "calculus"
|
||||
- "algebra"
|
||||
- "physics"
|
||||
- "chemistry"
|
||||
- "biology"
|
||||
- "history lesson"
|
||||
# Health & medical (non-travel)
|
||||
- "diagnosis"
|
||||
- "symptom"
|
||||
- "treatment"
|
||||
- "prescription"
|
||||
- "surgery"
|
||||
# Real estate
|
||||
- "real estate"
|
||||
- "mortgage"
|
||||
- "apartment"
|
||||
- "house price"
|
||||
# Dating & relationships
|
||||
- "dating"
|
||||
- "relationship advice"
|
||||
- "break up"
|
||||
- "tinder"
|
||||
# Gaming
|
||||
- "video game"
|
||||
- "gaming"
|
||||
- "playstation"
|
||||
- "xbox"
|
||||
- "fortnite"
|
||||
- "minecraft"
|
||||
|
||||
# CONTEXTUAL TRIGGERS
|
||||
# When combined with an identifier_word in the same sentence, triggers a block.
|
||||
additional_block_words:
|
||||
# Action/query words that confirm off-topic intent
|
||||
- "today"
|
||||
- "latest"
|
||||
- "score"
|
||||
- "won"
|
||||
- "winner"
|
||||
- "lost"
|
||||
- "write"
|
||||
- "build"
|
||||
- "create"
|
||||
- "develop"
|
||||
- "debug"
|
||||
- "fix"
|
||||
- "top"
|
||||
- "favorite"
|
||||
- "watch"
|
||||
- "listen"
|
||||
- "play"
|
||||
- "download"
|
||||
- "install"
|
||||
- "price"
|
||||
- "cost"
|
||||
- "buy"
|
||||
- "sell"
|
||||
- "vote"
|
||||
- "voted"
|
||||
- "opinion"
|
||||
- "who won"
|
||||
- "make"
|
||||
- "how to"
|
||||
- "tutorial"
|
||||
- "learn"
|
||||
- "teach"
|
||||
- "solve"
|
||||
- "calculate"
|
||||
- "convert"
|
||||
- "translate"
|
||||
|
||||
# ALWAYS BLOCK - Unambiguous off-topic phrases (blocked regardless of context)
|
||||
always_block_keywords:
|
||||
# News queries
|
||||
- keyword: "what's in the news"
|
||||
severity: "high"
|
||||
- keyword: "what is in the news"
|
||||
severity: "high"
|
||||
- keyword: "latest headlines"
|
||||
severity: "high"
|
||||
- keyword: "what happened in the world"
|
||||
severity: "high"
|
||||
# Jokes & fun
|
||||
- keyword: "tell me a joke"
|
||||
severity: "high"
|
||||
- keyword: "tell me a story"
|
||||
severity: "high"
|
||||
- keyword: "tell me a fun fact"
|
||||
severity: "high"
|
||||
- keyword: "tell me something interesting"
|
||||
severity: "high"
|
||||
# Coding requests
|
||||
- keyword: "write me code"
|
||||
severity: "high"
|
||||
- keyword: "write a script"
|
||||
severity: "high"
|
||||
- keyword: "write a program"
|
||||
severity: "high"
|
||||
- keyword: "help me code"
|
||||
severity: "high"
|
||||
- keyword: "fix my code"
|
||||
severity: "high"
|
||||
- keyword: "debug my code"
|
||||
severity: "high"
|
||||
# General knowledge
|
||||
- keyword: "capital of"
|
||||
severity: "high"
|
||||
- keyword: "who invented"
|
||||
severity: "high"
|
||||
- keyword: "how tall is"
|
||||
severity: "high"
|
||||
- keyword: "how old is"
|
||||
severity: "high"
|
||||
- keyword: "what year did"
|
||||
severity: "high"
|
||||
- keyword: "who is the president"
|
||||
severity: "high"
|
||||
# Math & homework
|
||||
- keyword: "solve this equation"
|
||||
severity: "high"
|
||||
- keyword: "what is 2+2"
|
||||
severity: "high"
|
||||
- keyword: "help me with my homework"
|
||||
severity: "high"
|
||||
# Recipes
|
||||
- keyword: "recipe for"
|
||||
severity: "high"
|
||||
- keyword: "how to cook"
|
||||
severity: "high"
|
||||
- keyword: "how to bake"
|
||||
severity: "high"
|
||||
# Relationship advice
|
||||
- keyword: "relationship advice"
|
||||
severity: "high"
|
||||
- keyword: "should I break up"
|
||||
severity: "high"
|
||||
- keyword: "dating advice"
|
||||
severity: "high"
|
||||
# AI / tech queries
|
||||
- keyword: "what is an AI gateway"
|
||||
severity: "high"
|
||||
- keyword: "explain machine learning"
|
||||
severity: "high"
|
||||
- keyword: "what is blockchain"
|
||||
severity: "high"
|
||||
- keyword: "what is cryptocurrency"
|
||||
severity: "high"
|
||||
|
||||
# EXCEPTIONS - Airline-adjacent contexts that should NOT be blocked
|
||||
exceptions:
|
||||
- "in-flight entertainment"
|
||||
- "flight entertainment"
|
||||
- "in-flight movie"
|
||||
- "airport news"
|
||||
- "travel news"
|
||||
- "airline news"
|
||||
- "flight news"
|
||||
- "aviation news"
|
||||
- "airport restaurant"
|
||||
- "airport lounge"
|
||||
- "travel recommend"
|
||||
- "destination recommend"
|
||||
- "flight price"
|
||||
- "ticket price"
|
||||
- "fare price"
|
||||
- "baggage cost"
|
||||
- "upgrade cost"
|
||||
- "booking cost"
|
||||
- "seat recommend"
|
||||
- "recommend seat"
|
||||
- "recommend flight"
|
||||
- "suggest flight"
|
||||
- "suggest seat"
|
||||
- "suggest upgrade"
|
||||
- "best seat"
|
||||
- "best flight"
|
||||
- "best fare"
|
||||
- "flight movie"
|
||||
- "explain my"
|
||||
- "explain the"
|
||||
- "explain flight"
|
||||
- "explain booking"
|
||||
|
|
@ -1,216 +0,0 @@
|
|||
"""
|
||||
Tests for the airline off-topic restriction policy template.
|
||||
|
||||
Verifies that off-topic messages are blocked and on-topic/conversational messages pass.
|
||||
"""
|
||||
|
||||
import os
|
||||
import sys
|
||||
|
||||
import pytest
|
||||
|
||||
sys.path.insert(
|
||||
0, os.path.abspath("../../")
|
||||
) # Adds the parent directory to the system path
|
||||
|
||||
from fastapi import HTTPException
|
||||
|
||||
from litellm.proxy.guardrails.guardrail_hooks.litellm_content_filter.content_filter import (
|
||||
ContentFilterGuardrail,
|
||||
)
|
||||
|
||||
POLICY_TEMPLATE_PATH = os.path.join(
|
||||
os.path.dirname(__file__),
|
||||
"../../../../../../litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/policy_templates/airline_off_topic_restriction.yaml",
|
||||
)
|
||||
|
||||
|
||||
def _make_guardrail():
|
||||
"""Create a ContentFilterGuardrail with the airline off-topic restriction loaded."""
|
||||
return ContentFilterGuardrail(
|
||||
guardrail_name="test-airline-off-topic",
|
||||
categories=[
|
||||
{
|
||||
"category": "airline_off_topic_restriction",
|
||||
"category_file": POLICY_TEMPLATE_PATH,
|
||||
"enabled": True,
|
||||
"action": "BLOCK",
|
||||
}
|
||||
],
|
||||
)
|
||||
|
||||
|
||||
class TestAirlineOffTopicRestriction:
|
||||
"""Test the airline off-topic restriction policy template."""
|
||||
|
||||
def test_on_topic_flight_booking(self):
|
||||
"""Airline booking questions should pass."""
|
||||
guardrail = _make_guardrail()
|
||||
# Should not raise
|
||||
result = guardrail._filter_single_text("I want to book a flight to Dubai")
|
||||
assert result == "I want to book a flight to Dubai"
|
||||
|
||||
def test_on_topic_baggage(self):
|
||||
"""Baggage questions should pass."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("What is the baggage allowance for economy?")
|
||||
assert result == "What is the baggage allowance for economy?"
|
||||
|
||||
def test_on_topic_checkin(self):
|
||||
"""Check-in questions should pass."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("How do I check in online?")
|
||||
assert "check in" in result
|
||||
|
||||
def test_on_topic_delay(self):
|
||||
"""Flight delay questions should pass."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("My flight is delayed, what are my options?")
|
||||
assert "delayed" in result
|
||||
|
||||
def test_conversational_hello(self):
|
||||
"""Greetings should pass."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("Hello")
|
||||
assert result == "Hello"
|
||||
|
||||
def test_conversational_thanks(self):
|
||||
"""Thank you should pass."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("Thank you for your help")
|
||||
assert "Thank you" in result
|
||||
|
||||
def test_conversational_help(self):
|
||||
"""Help requests should pass."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("Can you help me?")
|
||||
assert "help" in result
|
||||
|
||||
def test_conversational_yes_no(self):
|
||||
"""Simple yes/no should pass."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("Yes")
|
||||
assert result == "Yes"
|
||||
|
||||
def test_off_topic_news_always_block(self):
|
||||
"""News questions should be blocked via always_block_keywords."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("What's in the news today?")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_joke_always_block(self):
|
||||
"""Joke requests should be blocked via always_block_keywords."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("Tell me a joke")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_coding_always_block(self):
|
||||
"""Coding requests should be blocked via always_block_keywords."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("Write me code in python")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_ai_gateway_always_block(self):
|
||||
"""AI gateway questions should be blocked via always_block_keywords."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("What is an AI gateway?")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_capital_always_block(self):
|
||||
"""General knowledge questions should be blocked via always_block_keywords."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("What is the capital of France?")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_sports_conditional(self):
|
||||
"""Sports questions should be blocked via conditional matching."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("Who won the football game today?")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_recipe_always_block(self):
|
||||
"""Recipe questions should be blocked via always_block_keywords."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("Give me a recipe for pasta")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_movie_conditional(self):
|
||||
"""Movie questions with a block word should be blocked."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("What is the top movie to watch on Netflix?")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_on_topic_recommend_seat(self):
|
||||
"""Airline recommendation questions should pass (not false-positive)."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("Can you recommend the best seat?")
|
||||
assert "recommend" in result.lower()
|
||||
|
||||
def test_on_topic_explain_booking(self):
|
||||
"""Explain questions about airline topics should pass."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("Can you explain my booking details?")
|
||||
assert "explain" in result.lower()
|
||||
|
||||
def test_off_topic_stock_conditional(self):
|
||||
"""Stock market questions should be blocked via conditional matching."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("What is the stock price of Apple today?")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_homework_always_block(self):
|
||||
"""Homework requests should be blocked via always_block_keywords."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("Help me with my homework")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_relationship_always_block(self):
|
||||
"""Relationship advice should be blocked via always_block_keywords."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("Can you give me relationship advice?")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_exception_inflight_entertainment(self):
|
||||
"""In-flight entertainment questions should pass (exception)."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text(
|
||||
"What movies are available on the in-flight entertainment?"
|
||||
)
|
||||
assert "in-flight entertainment" in result.lower()
|
||||
|
||||
def test_exception_flight_price(self):
|
||||
"""Flight price questions should pass (exception)."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("What is the flight price to London?")
|
||||
assert "flight price" in result.lower()
|
||||
|
||||
def test_exception_travel_news(self):
|
||||
"""Travel news should pass (exception)."""
|
||||
guardrail = _make_guardrail()
|
||||
result = guardrail._filter_single_text("Any travel news I should know about?")
|
||||
assert "travel news" in result.lower()
|
||||
|
||||
def test_off_topic_president_always_block(self):
|
||||
"""Political questions should be blocked."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("Who is the president of the United States?")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
||||
def test_off_topic_blockchain_always_block(self):
|
||||
"""Blockchain questions should be blocked."""
|
||||
guardrail = _make_guardrail()
|
||||
with pytest.raises(HTTPException) as exc_info:
|
||||
guardrail._filter_single_text("What is blockchain technology?")
|
||||
assert exc_info.value.status_code == 403
|
||||
|
|
@ -532,55 +532,59 @@ export default function ComplianceUI({
|
|||
status: "pending",
|
||||
}));
|
||||
setTestResults(pendingResults);
|
||||
try {
|
||||
const { inputs, guardrail_errors } = await testPoliciesAndGuardrails(
|
||||
accessToken,
|
||||
{
|
||||
policy_names:
|
||||
selectedPolicies.length > 0 ? selectedPolicies : undefined,
|
||||
guardrail_names:
|
||||
selectedGuardrails.length > 0 ? selectedGuardrails : undefined,
|
||||
inputs: { texts: allTexts },
|
||||
request_data: {},
|
||||
input_type: "request",
|
||||
}
|
||||
);
|
||||
const actualResult: "blocked" | "allowed" =
|
||||
guardrail_errors.length > 0 ? "blocked" : "allowed";
|
||||
const triggeredBy =
|
||||
guardrail_errors.length > 0
|
||||
? guardrail_errors
|
||||
.map((e) => `${e.guardrail_name}: ${e.message}`)
|
||||
.join("; ")
|
||||
: undefined;
|
||||
const returnedTexts: (string | undefined)[] =
|
||||
Array.isArray(inputs?.texts) ? inputs.texts : [];
|
||||
setTestResults(
|
||||
pendingResults.map((row, index) => ({
|
||||
...row,
|
||||
// Send each text individually to get per-text blocked/allowed results.
|
||||
// Sending all texts in a single batch doesn't work because the guardrail
|
||||
// raises an HTTPException on the first blocked text, skipping the rest.
|
||||
const updatedResults = [...pendingResults];
|
||||
for (let i = 0; i < allTexts.length; i++) {
|
||||
try {
|
||||
const { inputs, guardrail_errors } = await testPoliciesAndGuardrails(
|
||||
accessToken,
|
||||
{
|
||||
policy_names:
|
||||
selectedPolicies.length > 0 ? selectedPolicies : undefined,
|
||||
guardrail_names:
|
||||
selectedGuardrails.length > 0 ? selectedGuardrails : undefined,
|
||||
inputs: { texts: [allTexts[i]] },
|
||||
request_data: {},
|
||||
input_type: "request",
|
||||
}
|
||||
);
|
||||
const actualResult: "blocked" | "allowed" =
|
||||
guardrail_errors.length > 0 ? "blocked" : "allowed";
|
||||
const triggeredBy =
|
||||
guardrail_errors.length > 0
|
||||
? guardrail_errors
|
||||
.map((e) => `${e.guardrail_name}: ${e.message}`)
|
||||
.join("; ")
|
||||
: undefined;
|
||||
const returnedText =
|
||||
Array.isArray(inputs?.texts) && inputs.texts.length > 0
|
||||
? inputs.texts[0]
|
||||
: undefined;
|
||||
updatedResults[i] = {
|
||||
...updatedResults[i],
|
||||
actualResult,
|
||||
isMatch:
|
||||
(row.expectedResult === "fail" && actualResult === "blocked") ||
|
||||
(row.expectedResult === "pass" && actualResult === "allowed"),
|
||||
(updatedResults[i].expectedResult === "fail" && actualResult === "blocked") ||
|
||||
(updatedResults[i].expectedResult === "pass" && actualResult === "allowed"),
|
||||
triggeredBy,
|
||||
returnedText: returnedTexts[index],
|
||||
returnedText,
|
||||
status: "complete" as const,
|
||||
}))
|
||||
);
|
||||
} catch (err) {
|
||||
const errorMessage = err instanceof Error ? err.message : String(err);
|
||||
setTestResults(
|
||||
pendingResults.map((row) => ({
|
||||
...row,
|
||||
};
|
||||
} catch (err) {
|
||||
const errorMessage = err instanceof Error ? err.message : String(err);
|
||||
updatedResults[i] = {
|
||||
...updatedResults[i],
|
||||
actualResult: "blocked" as const,
|
||||
isMatch: false,
|
||||
triggeredBy: `Error: ${errorMessage}`,
|
||||
status: "complete" as const,
|
||||
}))
|
||||
);
|
||||
} finally {
|
||||
setIsRunning(false);
|
||||
};
|
||||
}
|
||||
setTestResults([...updatedResults]);
|
||||
}
|
||||
setIsRunning(false);
|
||||
}, [
|
||||
accessToken,
|
||||
selectedPromptIds,
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue