This commit is contained in:
Timothy Jaeryang Baek 2026-10-26 18:54:49 +09:00
parent 8889e8d6af
commit 4a10c093ab
10 changed files with 94 additions and 25 deletions

View file

@ -3012,6 +3012,8 @@ DEFAULT_CONFIG = {
'rag.embedding_model': RAG_EMBEDDING_MODEL,
'rag.tokenizer_model': RAG_TOKENIZER_MODEL,
'rag.embedding_batch_size': RAG_EMBEDDING_BATCH_SIZE,
'rag.embedding_query_prefix': RAG_EMBEDDING_QUERY_PREFIX,
'rag.embedding_content_prefix': RAG_EMBEDDING_CONTENT_PREFIX,
'rag.enable_async_embedding': ENABLE_ASYNC_EMBEDDING,
'rag.embedding_concurrent_requests': RAG_EMBEDDING_CONCURRENT_REQUESTS,
'rag.reranking_engine': RAG_RERANKING_ENGINE,

View file

@ -4,7 +4,6 @@ import re
import time
from typing import Any, Optional
from open_webui.config import RAG_EMBEDDING_QUERY_PREFIX
from open_webui.models.config import Config
from open_webui.models.knowledge import KnowledgeModel
@ -105,7 +104,7 @@ async def _retrieve_qdrant(connection, auth_config, knowledge, query, count, emb
source_config = _source_config(knowledge)
vector_field = source_config.get('vector_field') or None
vector = await embedding_function(query, prefix=RAG_EMBEDDING_QUERY_PREFIX)
vector = await embedding_function(query, prefix=await Config.get('rag.embedding_query_prefix'))
def _search():
client = QdrantClient(
@ -154,7 +153,7 @@ async def _retrieve_milvus(connection, auth_config, knowledge, query, count, emb
content_field = source_config.get('content_field') or 'data.text'
metadata_field = source_config.get('metadata_field') or 'metadata'
vector = await embedding_function(query, prefix=RAG_EMBEDDING_QUERY_PREFIX)
vector = await embedding_function(query, prefix=await Config.get('rag.embedding_query_prefix'))
def _search():
client_kwargs = {
@ -232,7 +231,7 @@ async def _retrieve_pgvector(connection, auth_config, knowledge, query, count, e
metadata_field = source_config.get('metadata_field') or 'vmetadata'
document_id_field = source_config.get('document_id_field') or 'id'
vector = await embedding_function(query, prefix=RAG_EMBEDDING_QUERY_PREFIX)
vector = await embedding_function(query, prefix=await Config.get('rag.embedding_query_prefix'))
def _search():
from psycopg import sql

View file

@ -19,9 +19,7 @@ from langchain_classic.retrievers import (
)
from langchain_core.documents import Document
from open_webui.config import (
RAG_EMBEDDING_CONTENT_PREFIX,
RAG_EMBEDDING_PREFIX_FIELD_NAME,
RAG_EMBEDDING_QUERY_PREFIX,
VECTOR_DB,
)
from open_webui.constants import ERROR_MESSAGES
@ -321,7 +319,7 @@ class VectorSearchRetriever(BaseRetriever):
*,
run_manager: CallbackManagerForRetrieverRun,
) -> list[Document]:
embedding = await self.embedding_function(query, RAG_EMBEDDING_QUERY_PREFIX)
embedding = await self.embedding_function(query, await Config.get('rag.embedding_query_prefix'))
result = await ASYNC_VECTOR_DB_CLIENT.search(
collection_name=self.collection_name,
vectors=[embedding],
@ -433,7 +431,7 @@ async def query_doc_with_native_hybrid_search(
query_vectors = []
if hybrid_bm25_weight < 1:
query_vectors = [await embedding_function(query, RAG_EMBEDDING_QUERY_PREFIX)]
query_vectors = [await embedding_function(query, await Config.get('rag.embedding_query_prefix'))]
result = await ASYNC_VECTOR_DB_CLIENT.hybrid_search(
collection_name=collection_name,
@ -696,6 +694,7 @@ async def query_collection(
) -> dict:
config = await Config.get_many(
'rag.enable_hybrid_search',
'rag.embedding_query_prefix',
'rag.top_k_reranker',
'rag.relevance_threshold',
'rag.hybrid_bm25_weight',
@ -749,7 +748,7 @@ async def query_collection(
return {'distances': [[]], 'documents': [[]], 'metadatas': [[]]}
# Generate all query embeddings (in one call)
query_embeddings = await embedding_function(queries, prefix=RAG_EMBEDDING_QUERY_PREFIX)
query_embeddings = await embedding_function(queries, prefix=config['rag.embedding_query_prefix'])
log.debug('query_collection: processing %s queries across %s collections', len(queries), len(collection_names))
task_results = await asyncio.gather(
@ -1851,9 +1850,11 @@ class RerankCompressor(BaseDocumentCompressor):
if reranking:
scores = await asyncio.to_thread(self.reranking_function, query, documents)
else:
query_embedding = await self.embedding_function(query, RAG_EMBEDDING_QUERY_PREFIX)
query_embedding = await self.embedding_function(query, await Config.get('rag.embedding_query_prefix'))
doc_texts = [doc.page_content for doc in documents]
document_embedding = await self.embedding_function(doc_texts, RAG_EMBEDDING_CONTENT_PREFIX)
document_embedding = await self.embedding_function(
doc_texts, await Config.get('rag.embedding_content_prefix')
)
scores = cosine_similarity(query_embedding, document_embedding)
if scores is not None:

View file

@ -14,7 +14,6 @@ from fastapi.responses import StreamingResponse
from open_webui.config import (
BYPASS_ADMIN_ACCESS_CONTROL,
ENABLE_KNOWLEDGE_FILE_RETENTION,
RAG_EMBEDDING_CONTENT_PREFIX,
)
from open_webui.constants import ERROR_MESSAGES
from open_webui.events import EVENTS, publish_event
@ -98,7 +97,9 @@ async def embed_knowledge_base_metadata(
"""Generate and store embedding for knowledge base."""
try:
content = f'{name}\n\n{description}' if description else name
embedding = await request.app.state.EMBEDDING_FUNCTION(content, prefix=RAG_EMBEDDING_CONTENT_PREFIX)
embedding = await request.app.state.EMBEDDING_FUNCTION(
content, prefix=await Config.get('rag.embedding_content_prefix')
)
await ASYNC_VECTOR_DB_CLIENT.upsert(
collection_name=KNOWLEDGE_BASES_COLLECTION,
items=[

View file

@ -4,7 +4,6 @@ import logging
from typing import Literal
from fastapi import APIRouter, Depends, HTTPException, Request, status
from open_webui.config import RAG_EMBEDDING_CONTENT_PREFIX, RAG_EMBEDDING_QUERY_PREFIX
from open_webui.constants import ERROR_MESSAGES
from open_webui.events import EVENTS, publish_event
from open_webui.internal.db import get_async_session
@ -144,7 +143,7 @@ async def reindex_memory_vectors_for_user(
# One list call so batching and the concurrency limit apply.
vectors = await request.app.state.EMBEDDING_FUNCTION(
[memory_vector_text(memory.content, memory.path) for memory in memories],
prefix=RAG_EMBEDDING_CONTENT_PREFIX,
prefix=await Config.get('rag.embedding_content_prefix'),
user=user,
)
@ -200,7 +199,9 @@ async def add_memory(
)
vector = await request.app.state.EMBEDDING_FUNCTION(
memory_vector_text(memory.content, memory.path), prefix=RAG_EMBEDDING_CONTENT_PREFIX, user=user
memory_vector_text(memory.content, memory.path),
prefix=await Config.get('rag.embedding_content_prefix'),
user=user,
)
await upsert_memory_vectors_or_reindex(
@ -256,6 +257,7 @@ async def update_memories(
delete_ids = []
response = []
content_prefix = await Config.get('rag.embedding_content_prefix')
for result in results:
memory = result.get('memory')
if isinstance(memory, MemoryModel):
@ -263,7 +265,7 @@ async def update_memories(
if result.get('status') in {'created', 'updated'}:
vector = await request.app.state.EMBEDDING_FUNCTION(
memory_vector_text(memory.content, memory.path),
prefix=RAG_EMBEDDING_CONTENT_PREFIX,
prefix=content_prefix,
user=user,
)
upsert_items.append(
@ -340,7 +342,9 @@ async def query_memory(
if not memories:
raise HTTPException(status_code=404, detail='No memories found for user')
vector = await request.app.state.EMBEDDING_FUNCTION(form_data.content, prefix=RAG_EMBEDDING_QUERY_PREFIX, user=user)
vector = await request.app.state.EMBEDDING_FUNCTION(
form_data.content, prefix=await Config.get('rag.embedding_query_prefix'), user=user
)
results = await ASYNC_VECTOR_DB_CLIENT.search(
collection_name=f'user-memory-{user.id}',
@ -570,7 +574,9 @@ async def update_memory_by_id(
if form_data.content is not None or form_data.path is not None:
vector = await request.app.state.EMBEDDING_FUNCTION(
memory_vector_text(memory.content, memory.path), prefix=RAG_EMBEDDING_CONTENT_PREFIX, user=user
memory_vector_text(memory.content, memory.path),
prefix=await Config.get('rag.embedding_content_prefix'),
user=user,
)
await upsert_memory_vectors_or_reindex(

View file

@ -37,10 +37,8 @@ from langchain_text_splitters import (
from open_webui.config import (
Config,
DEFAULT_LOCALE,
RAG_EMBEDDING_CONTENT_PREFIX,
RAG_EMBEDDING_MODEL_AUTO_UPDATE,
RAG_EMBEDDING_MODEL_TRUST_REMOTE_CODE,
RAG_EMBEDDING_QUERY_PREFIX,
RAG_RERANKING_MODEL_AUTO_UPDATE,
RAG_RERANKING_MODEL_TRUST_REMOTE_CODE,
UPLOAD_DIR,
@ -347,6 +345,8 @@ RETRIEVAL_CONFIG_KEYS = {
'RAG_EMBEDDING_BATCH_SIZE': 'rag.embedding_batch_size',
'RAG_EMBEDDING_CONCURRENT_REQUESTS': 'rag.embedding_concurrent_requests',
'RAG_EMBEDDING_ENGINE': 'rag.embedding_engine',
'RAG_EMBEDDING_QUERY_PREFIX': 'rag.embedding_query_prefix',
'RAG_EMBEDDING_CONTENT_PREFIX': 'rag.embedding_content_prefix',
'RAG_EMBEDDING_MODEL': 'rag.embedding_model',
'RAG_TOKENIZER_MODEL': 'rag.tokenizer_model',
'RAG_EXTERNAL_RERANKER_API_KEY': 'rag.external_reranker_api_key',
@ -443,6 +443,8 @@ async def get_embedding_config(request: Request, user=Depends(get_admin_user)):
return {
'status': True,
'RAG_EMBEDDING_ENGINE': config['rag.embedding_engine'],
'RAG_EMBEDDING_QUERY_PREFIX': config['rag.embedding_query_prefix'],
'RAG_EMBEDDING_CONTENT_PREFIX': config['rag.embedding_content_prefix'],
'RAG_EMBEDDING_MODEL': config['rag.embedding_model'],
'RAG_EMBEDDING_BATCH_SIZE': config['rag.embedding_batch_size'],
'ENABLE_ASYNC_EMBEDDING': config['rag.enable_async_embedding'],
@ -484,6 +486,8 @@ class EmbeddingModelUpdateForm(BaseModel):
ollama_config: OllamaConfigForm | None = None
azure_openai_config: AzureOpenAIConfigForm | None = None
RAG_EMBEDDING_ENGINE: str
RAG_EMBEDDING_QUERY_PREFIX: str | None = None
RAG_EMBEDDING_CONTENT_PREFIX: str | None = None
RAG_EMBEDDING_MODEL: str
RAG_EMBEDDING_BATCH_SIZE: int | None = 1
ENABLE_ASYNC_EMBEDDING: bool | None = True
@ -521,6 +525,10 @@ async def update_embedding_config(request: Request, form_data: EmbeddingModelUpd
'rag.enable_async_embedding': form_data.ENABLE_ASYNC_EMBEDDING,
'rag.embedding_concurrent_requests': form_data.RAG_EMBEDDING_CONCURRENT_REQUESTS,
}
if 'RAG_EMBEDDING_QUERY_PREFIX' in form_data.model_fields_set:
updates['rag.embedding_query_prefix'] = form_data.RAG_EMBEDDING_QUERY_PREFIX or None
if 'RAG_EMBEDDING_CONTENT_PREFIX' in form_data.model_fields_set:
updates['rag.embedding_content_prefix'] = form_data.RAG_EMBEDDING_CONTENT_PREFIX or None
if form_data.RAG_EMBEDDING_ENGINE == 'openai' and form_data.openai_config is not None:
updates['rag.openai.api_base_url'] = form_data.openai_config.url or ''
updates['rag.openai.api_key'] = form_data.openai_config.key or ''
@ -572,6 +580,8 @@ async def update_embedding_config(request: Request, form_data: EmbeddingModelUpd
return {
'status': True,
'RAG_EMBEDDING_ENGINE': config['rag.embedding_engine'],
'RAG_EMBEDDING_QUERY_PREFIX': config['rag.embedding_query_prefix'],
'RAG_EMBEDDING_CONTENT_PREFIX': config['rag.embedding_content_prefix'],
'RAG_EMBEDDING_MODEL': config['rag.embedding_model'],
'RAG_EMBEDDING_BATCH_SIZE': config['rag.embedding_batch_size'],
'ENABLE_ASYNC_EMBEDDING': config['rag.enable_async_embedding'],
@ -1526,7 +1536,7 @@ def save_docs_to_vector_db(
future = asyncio.run_coroutine_threadsafe(
embedding_function(
list(map(lambda x: x.replace('\n', ' '), texts)),
prefix=RAG_EMBEDDING_CONTENT_PREFIX,
prefix=config['rag.embedding_content_prefix'],
user=user,
),
request.app.state.main_loop,
@ -2790,7 +2800,7 @@ async def query_doc_handler(
)
else:
query_embedding = await request.app.state.EMBEDDING_FUNCTION(
form_data.query, prefix=RAG_EMBEDDING_QUERY_PREFIX, user=user
form_data.query, prefix=config['rag.embedding_query_prefix'], user=user
)
# query_doc wraps a blocking get_vector_db_client().search call;
# offload so the request's event loop stays responsive.

View file

@ -14,7 +14,6 @@ from urllib.parse import unquote
from fastapi import HTTPException, Request
from open_webui.config import RAG_EMBEDDING_QUERY_PREFIX
from open_webui.env import (
KNOWLEDGE_GREP_MAX_MATCHES,
VIEW_FILE_DEFAULT_MAX_CHARS,
@ -3496,7 +3495,9 @@ async def query_knowledge_bases(
if not embedding_function:
return JSONCodec.dumps({'error': 'Embedding function not configured'})
user_model = UserModel(**__user__)
query_embedding = await embedding_function(query, prefix=RAG_EMBEDDING_QUERY_PREFIX, user=user_model)
query_embedding = await embedding_function(
query, prefix=await Config.get('rag.embedding_query_prefix'), user=user_model
)
# Min-heap of (distance, knowledge_base_id) - only holds top `count` results
top_results_heap = []

View file

@ -203,6 +203,8 @@ type EmbeddingModelUpdateForm = {
azure_openai_config?: AzureOpenAIConfigForm;
RAG_EMBEDDING_ENGINE: string;
RAG_EMBEDDING_MODEL: string;
RAG_EMBEDDING_QUERY_PREFIX?: string;
RAG_EMBEDDING_CONTENT_PREFIX?: string;
RAG_EMBEDDING_BATCH_SIZE?: number;
ENABLE_ASYNC_EMBEDDING?: boolean;
RAG_EMBEDDING_CONCURRENT_REQUESTS?: number;

View file

@ -31,6 +31,7 @@
import Textarea from '$lib/components/common/Textarea.svelte';
import Spinner from '$lib/components/common/Spinner.svelte';
import SettingsSelect from '$lib/components/common/SettingsSelect.svelte';
import Collapsible from '$lib/components/common/Collapsible.svelte';
import AdminSettingField from './AdminSettingField.svelte';
import AdminSettingRow from './AdminSettingRow.svelte';
import AdminSettingSection from './AdminSettingSection.svelte';
@ -43,10 +44,13 @@
let showResetConfirm = false;
let showResetUploadDirConfirm = false;
let showReindexConfirm = false;
let showEmbeddingAdvanced = false;
let showExternalDocumentLoaderHeadersHint = false;
let RAG_EMBEDDING_ENGINE = '';
let RAG_EMBEDDING_MODEL = '';
let RAG_EMBEDDING_QUERY_PREFIX = '';
let RAG_EMBEDDING_CONTENT_PREFIX = '';
let RAG_EMBEDDING_BATCH_SIZE = 1;
let ENABLE_ASYNC_EMBEDDING = true;
let RAG_EMBEDDING_CONCURRENT_REQUESTS = 0;
@ -126,6 +130,8 @@
const payload: Parameters<typeof updateEmbeddingConfig>[1] = {
RAG_EMBEDDING_ENGINE: RAG_EMBEDDING_ENGINE,
RAG_EMBEDDING_MODEL: RAG_EMBEDDING_MODEL,
RAG_EMBEDDING_QUERY_PREFIX,
RAG_EMBEDDING_CONTENT_PREFIX,
RAG_EMBEDDING_BATCH_SIZE: RAG_EMBEDDING_BATCH_SIZE,
ENABLE_ASYNC_EMBEDDING: ENABLE_ASYNC_EMBEDDING,
RAG_EMBEDDING_CONCURRENT_REQUESTS: RAG_EMBEDDING_CONCURRENT_REQUESTS
@ -301,6 +307,9 @@
if (embeddingConfig) {
RAG_EMBEDDING_ENGINE = embeddingConfig.RAG_EMBEDDING_ENGINE;
RAG_EMBEDDING_MODEL = embeddingConfig.RAG_EMBEDDING_MODEL;
RAG_EMBEDDING_QUERY_PREFIX = embeddingConfig.RAG_EMBEDDING_QUERY_PREFIX ?? '';
RAG_EMBEDDING_CONTENT_PREFIX = embeddingConfig.RAG_EMBEDDING_CONTENT_PREFIX ?? '';
showEmbeddingAdvanced = Boolean(RAG_EMBEDDING_QUERY_PREFIX || RAG_EMBEDDING_CONTENT_PREFIX);
RAG_EMBEDDING_BATCH_SIZE = embeddingConfig.RAG_EMBEDDING_BATCH_SIZE ?? 1;
ENABLE_ASYNC_EMBEDDING = embeddingConfig.ENABLE_ASYNC_EMBEDDING ?? true;
RAG_EMBEDDING_CONCURRENT_REQUESTS = embeddingConfig.RAG_EMBEDDING_CONCURRENT_REQUESTS ?? 0;
@ -1225,6 +1234,40 @@
/>
</AdminSettingRow>
{/if}
<Collapsible
title={$i18n.t('Advanced')}
bind:open={showEmbeddingAdvanced}
buttonClassName="w-full py-1 text-xs text-gray-600 hover:text-gray-900 dark:text-gray-400 dark:hover:text-white transition-colors"
>
<div slot="content" class="flex flex-col gap-2.5 pt-2.5">
<AdminSettingField
label={$i18n.t('settings.admin.documents.embeddingQueryPrefix.label')}
description={$i18n.t('settings.admin.documents.embeddingQueryPrefix.description')}
forId="embedding-query-prefix"
>
<input
id="embedding-query-prefix"
class={inputClass}
placeholder={$i18n.t('Optional')}
bind:value={RAG_EMBEDDING_QUERY_PREFIX}
/>
</AdminSettingField>
<AdminSettingField
label={$i18n.t('settings.admin.documents.embeddingContentPrefix.label')}
description={$i18n.t('settings.admin.documents.embeddingContentPrefix.description')}
forId="embedding-content-prefix"
>
<input
id="embedding-content-prefix"
class={inputClass}
placeholder={$i18n.t('Optional')}
bind:value={RAG_EMBEDDING_CONTENT_PREFIX}
/>
</AdminSettingField>
</div>
</Collapsible>
</AdminSettingSection>
{/if}

View file

@ -2777,6 +2777,10 @@
"settings.admin.documents.documentIntelligenceModel.label": "Document Intelligence Model",
"settings.admin.documents.documentLoaderUrl.description": "External service endpoint used to load document content.",
"settings.admin.documents.documentLoaderUrl.label": "Document Loader URL",
"settings.admin.documents.embeddingQueryPrefix.label": "Query Prefix",
"settings.admin.documents.embeddingQueryPrefix.description": "Prepended to search queries before embedding. Include any required trailing spaces.",
"settings.admin.documents.embeddingContentPrefix.label": "Content Prefix",
"settings.admin.documents.embeddingContentPrefix.description": "Prepended to content before embedding. Include any required trailing spaces. Reindex existing data after changing this.",
"settings.admin.documents.embeddingBatchSize.description": "Number of items processed per embedding batch.",
"settings.admin.documents.embeddingBatchSize.label": "Embedding Batch Size",
"settings.admin.documents.embeddingConcurrentRequests.description": "Maximum concurrent embedding requests. Set to 0 for unlimited.",