ReMe/reme2/memory/ingestor.py
huangsen b72690d954 feat: add Neo4j backend and refactor file graph architecture
- add neo4j dependency to project requirements
- introduce NetworkXFileGraph to replace LocalFileGraph implementation
- rename local_file_graph.py to networkx_file_graph.py with updated
  component registration as 'networkx'
- remove pickle persistence logic from NetworkX backend, simplify
  initialization
- update Neo4jFileGraph to return FileLink objects instead of
  (FileNode, FileLink) tuples from get_inlinks/get_outlinks methods
- remove unused AsyncIterator import and adjust method signatures
- add BareFileParser for handling binary/attachment files without
  content parsing
- move wikilink resolution utilities to dedicated utility module
- refactor memory I/O to use file graph's link resolution methods
  directly
- remove link extraction utilities from schema module, consolidate
  in utils.wikilink_resolver
2026-05-14 19:46:18 +08:00

217 lines
8.3 KiB
Python

"""Smart Ingestor — ReAct agent over the Memory File System.
Per `structure.md` L31-44, the Ingestor is the SSOT engine: the **single
write entry point** to the markdown vault. Every mutation (create, body
edit, frontmatter flip, rename, delete, archive) flows through here.
Mirrors `Summarizer`'s pattern — drives a `ReActAgent` whose toolkit is
built by `agent_toolkit.build_agent_toolkit`. The agent runs its own
R-M-W loop: read related files via tools, decide which ones to mutate,
call the right write tool. Every write tool records into an audit list,
so the caller gets a deterministic mutation trail regardless of how the
agent's reasoning unfolded.
When no LLM is configured, falls back to a direct create from
`target_path` + `metadata` + `content`. Edits/renames/deletes are not
available without an LLM.
"""
from __future__ import annotations
import datetime
import json
import zoneinfo
from pathlib import Path
from agentscope.agent import ReActAgent
from agentscope.message import Msg
from agentscope.tool import Toolkit
from pydantic import BaseModel, Field
from .runtime_response import _set_answer, _to_jsonable
from . import memory_io
from .memory_io import create_file
from .agent_toolkit import build_agent_toolkit
from ..component import R
from ..component.base_step import BaseStep
from ..enumeration import ComponentEnum
from ..utils.wikilink_resolver import extract_wikilinks
class IngestResult(BaseModel):
"""Audit trail for a single ingest call."""
applied: list[dict] = Field(default_factory=list, description="Successful ops with paths.")
rejected: list[dict] = Field(default_factory=list, description="Ops the validator refused.")
failed: list[dict] = Field(default_factory=list, description="Ops that errored at apply time.")
skipped: bool = Field(default=False, description="True if the LLM returned a SkipOp.")
used_llm: bool = Field(default=False, description="False = degraded path (no LLM configured).")
@property
def success(self) -> bool:
"""Skipped or any-applied with no failures = success."""
if self.skipped:
return True
return len(self.applied) > 0 and len(self.failed) == 0
@R.register("ingestor")
class Ingestor(BaseStep):
"""R-M-W ingestor exposed as a single-write-entry MCP tool.
Inputs (read from RuntimeContext):
content (str, required): the material being ingested.
hint (str, optional): caller guidance to the LLM.
target_path (str, optional): suggested file path; required for
the degraded path (no LLM).
metadata (dict, optional): suggested frontmatter; used by
the degraded path or as a hint to the LLM.
related_paths (list[str], optional): explicit related files,
auto-extended with wikilinks parsed from `content`.
Output (written to context.response.answer):
IngestResult JSON — applied/failed lists plus used_llm flag and
the agent's final-message summary.
"""
def __init__(
self,
toolkit: Toolkit | None = None,
console_enabled: bool = False,
timezone: str | None = None,
**kwargs,
):
super().__init__(**kwargs)
self.toolkit = toolkit
self.console_enabled = console_enabled
self.timezone = timezone
self._protocol = (Path(__file__).parent / "protocol.md").read_text(encoding="utf-8")
def _now(self) -> datetime.datetime:
if self.timezone:
try:
return datetime.datetime.now(zoneinfo.ZoneInfo(self.timezone))
except Exception as e:
self.logger.error(f"Invalid timezone: {self.timezone}, error={e}")
return datetime.datetime.now()
def _working_dir(self) -> Path:
vr = getattr(self.file_store, "working_dir", None)
return Path(vr).resolve() if vr else Path.cwd().resolve()
async def execute(self):
assert self.context is not None
mode: str = (self.context.get("mode") or "distill").lower()
if mode == "log":
await self._delegate_to_sync()
return
if mode != "distill":
self.context.response.success = False
_set_answer(
self.context,
{
"error": f"unknown mode {mode!r}; expected 'log' or 'distill'",
},
)
return
content: str = self.context.get("content", "") or ""
hint: str = self.context.get("hint", "") or ""
target_path: str = self.context.get("target_path") or ""
metadata: dict = dict(self.context.get("metadata") or {})
related_paths: list[str] = list(self.context.get("related_paths") or [])
assert content, "content is required"
# Auto-discover wikilink targets in content as a hint for the agent.
for link in extract_wikilinks(content):
hit = memory_io.resolve_wikilink(self.file_store, link)["path"]
if hit and hit not in related_paths:
related_paths.append(hit)
as_llm = self._get_component_optional(ComponentEnum.AS_LLM, "default", "model")
if as_llm is None:
result = self._degraded(target_path, metadata, content)
self.context.response.success = result.success
_set_answer(self.context, result.model_dump())
return
working_dir = self._working_dir()
audit: list[dict] = []
toolkit = build_agent_toolkit(self.app_context, audit=audit, toolkit=self.toolkit)
agent = ReActAgent(
name="reme_ingestor",
model=self.as_llm,
sys_prompt=self.prompt_format(
"system_prompt",
working_dir=str(working_dir),
protocol=self._protocol,
),
formatter=self.as_llm_formatter,
toolkit=toolkit,
)
agent.set_console_output_enabled(self.console_enabled)
user_message: str = self.prompt_format(
"user_message",
today=self._now().strftime("%Y-%m-%d"),
working_dir=str(working_dir),
hint=hint or "(none)",
target_path=target_path or "(none)",
metadata=json.dumps(_to_jsonable(metadata), ensure_ascii=False),
related=json.dumps(related_paths, ensure_ascii=False),
content=content,
)
final_msg: Msg = await agent.reply(
Msg(name="reme", role="user", content=user_message),
)
summary = final_msg.get_text_content() or ""
result = IngestResult(used_llm=True)
for entry in audit:
(result.applied if entry.get("ok") else result.failed).append(entry)
if not audit and summary.strip().upper().startswith("SKIP"):
result.skipped = True
self.context.response.success = result.success
payload = result.model_dump()
payload["agent_summary"] = summary
_set_answer(self.context, payload)
def _degraded(self, target_path: str, metadata: dict, content: str) -> IngestResult:
"""Without an LLM, only direct create from explicit target_path is
supported. Useful for tests and bootstrap scripts."""
result = IngestResult(used_llm=False)
if not target_path:
result.failed.append(
{
"op": "create",
"ok": False,
"error": "degraded path: target_path is required when no LLM is configured",
}
)
return result
path = Path(target_path)
if not path.is_absolute():
path = self._working_dir() / path
ok, payload = create_file(
self.file_store,
path,
metadata=metadata,
content=content,
)
bucket = result.applied if ok else result.failed
bucket.append({"op": "create", "ok": ok, "path": str(path), "result": payload})
return result
async def _delegate_to_sync(self) -> None:
"""Hot-path event-folder upsert — same code path as the standalone
`sync` step. Lazy-instantiated so we don't pay the construction
cost on every distill call."""
from .sync import Sync
if getattr(self, "_sync_step", None) is None:
self._sync_step = Sync(app_context=self.app_context)
await self._sync_step(self.context)