mirror of
https://github.com/tiennm99/DocsGPT.git
synced 2026-10-15 20:15:30 +00:00
Alex
574f96341e
The backend import package is now docsgpt, the name it will carry on PyPI; application was far too generic to install into anyone's site-packages. git mv plus a mechanical rewrite of every import, dotted string and path reference: 734 Python files, the compose files, Dockerfile, workflows, docs, setup scripts, devcontainer, k8s manifests, vscode config, pytest and coverage config, .gitignore. Behaviour is unchanged. Kept for one release: - A top-level application package whose meta-path finder resolves application.x.y to the already-imported docsgpt.x.y object, so old imports and entry points (celery -A application.app.celery, uvicorn application.asgi:asgi_app) keep working with a FutureWarning. - Celery registers every application.* task name as an alias of its docsgpt.* task on start-up, so messages queued by the previous release still run. The redbeat key prefix moves to redbeat:docsgpt:v2: so schedule entries the previous release wrote are left unread instead of firing twice. The backend image builds from the repository root (docker build -f docsgpt/Dockerfile .) so it can ship the alias package; a root .dockerignore allow-lists docsgpt/ and application/ and keeps caches, local data, .env files, the sample index files and the Dockerfile out. Compose and the image workflows point at the new context.
86 lines
3.3 KiB
Python
86 lines
3.3 KiB
Python
"""Hybrid retriever fusing vector and keyword search via RRF.
|
|
|
|
Subclasses :class:`ClassicRAG` so the Dispatcher builds it with identical
|
|
ctor kwargs and it inherits rephrase + token-budgeting. Only the per-source
|
|
fetch is overridden: for each vector store it pulls vector hits and keyword
|
|
hits, then fuses them with Reciprocal Rank Fusion. Stores without keyword
|
|
support (``keyword_search`` returns ``[]``) reduce to exact vector-only
|
|
behaviour.
|
|
"""
|
|
|
|
from typing import Any, List, Optional
|
|
|
|
from docsgpt.retriever.classic_rag import ClassicRAG
|
|
|
|
RRF_K = 60
|
|
|
|
|
|
def _doc_key(doc):
|
|
"""Stable identity for a hit so the same chunk fuses across both lists."""
|
|
if hasattr(doc, "page_content") and hasattr(doc, "metadata"):
|
|
content = doc.page_content
|
|
metadata = doc.metadata or {}
|
|
else:
|
|
content = doc.get("text", doc.get("page_content", ""))
|
|
metadata = doc.get("metadata") or {}
|
|
source = metadata.get("source", "")
|
|
return (source, content)
|
|
|
|
|
|
def fuse_with_scores(vector_hits, keyword_hits, k=RRF_K):
|
|
"""Fuse two ranked hit lists by RRF, keeping each hit's fused score.
|
|
|
|
Each list contributes ``1 / (k + rank)`` per document (rank 0-based);
|
|
documents are returned as ``(doc, fused_score)`` ordered by score, highest
|
|
first. A document present in only one list is ranked solely on that list's
|
|
contribution, so an empty ``keyword_hits`` yields exactly the vector
|
|
ordering.
|
|
"""
|
|
scores = {}
|
|
docs = {}
|
|
for hits in (vector_hits, keyword_hits):
|
|
for rank, doc in enumerate(hits):
|
|
key = _doc_key(doc)
|
|
scores[key] = scores.get(key, 0.0) + 1.0 / (k + rank)
|
|
if key not in docs:
|
|
docs[key] = doc
|
|
ordered = sorted(docs.keys(), key=lambda key: scores[key], reverse=True)
|
|
return [(docs[key], scores[key]) for key in ordered]
|
|
|
|
|
|
class HybridRetriever(ClassicRAG):
|
|
"""ClassicRAG variant that fuses vector + keyword search with RRF."""
|
|
|
|
def _score_kind(self, docsearch: Any) -> str:
|
|
"""RRF scores rank hits against each other, not against a similarity
|
|
cutoff — they are not comparable to the store's cosine scores, so they
|
|
carry their own label."""
|
|
return "rrf"
|
|
|
|
def _fetch_candidates(
|
|
self,
|
|
docsearch: Any,
|
|
question: str,
|
|
src_k: int,
|
|
score_threshold: Optional[float],
|
|
query_vector: Optional[List[float]] = None,
|
|
) -> List[Any]:
|
|
"""Return RRF-fused vector+keyword hits for one vector store.
|
|
|
|
Inherits the per-source resolution and budgeting from
|
|
:meth:`ClassicRAG._get_data`; only candidate sourcing differs.
|
|
RRF scores are not cosine similarities, so ``score_threshold`` is
|
|
intentionally not applied to the fused list. ``query_vector`` is the
|
|
retrieval's single query embedding — the keyword half never needs one.
|
|
"""
|
|
candidate_k = min(max(src_k * 2, 20), 500)
|
|
vector_kwargs = {"k": candidate_k}
|
|
if query_vector is not None:
|
|
vector_kwargs["query_vector"] = query_vector
|
|
vector_hits = docsearch.search(question, **vector_kwargs)
|
|
keyword_hits = docsearch.keyword_search(question, k=candidate_k)
|
|
fused = fuse_with_scores(vector_hits, keyword_hits)
|
|
if self.include_scores:
|
|
return fused
|
|
return [doc for doc, _ in fused]
|