orbitalo/homelab-brain
1
0
Fork 0
Code Issues 37 Pull requests Projects Releases Packages Wiki Activity Actions

fix(rag): RRF fusion, 512-char snippets, 15 candidates — speed+quality

Browse source
This commit is contained in:
Homelab Cursor 2026-03-28 16:51:06 +01:00
parent 3c455e7ad7
commit a3735bf265
1 changed files with 26 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

36
homelab-ai-bot/tools/rag.py
View file

@ -22,8 +22,9 @@ EMBED_MODEL = "nomic-embed-text"
# Cross-Encoder Reranking (CT 123, pve-hetzner LAN)
RERANKER_URL = "http://10.10.10.123:8099"
RERANK_CANDIDATES = 30
RERANK_TIMEOUT = 120
RERANK_CANDIDATES = 15
RERANK_TIMEOUT = 30
RERANK_SNIPPET_CHARS = 512
MIN_TOP_K = 5
# Breite Übersichten: mehr ES-Runden, mehr distinct Treffer (pro vollem Pfad docnm_kwd)
@ -207,12 +208,14 @@ def _es_hybrid_search(query: str, es_size: int) -> dict:
def _snippet_for_rerank(src: dict) -> str:
doc_name = src.get("docnm_kwd") or ""
raw = src.get("content_with_weight") or src.get("content_de") or ""
return raw[:4000]
prefix = doc_name[:120] + "\n" if doc_name else ""
return prefix + raw[:RERANK_SNIPPET_CHARS]
def _rerank_hits(query: str, hits: list) -> tuple[list, bool]:
"""Sortiert die ersten RERANK_CANDIDATES Treffer per Cross-Encoder neu."""
"""Rerankt mit Cross-Encoder, kombiniert Score mit ES-Rang (RRF)."""
if not hits or not RERANKER_URL:
return hits, False
to_score = hits[:RERANK_CANDIDATES]
@ -241,17 +244,27 @@ def _rerank_hits(query: str, hits: list) -> tuple[list, bool]:
len(to_score),
)
return hits, False
indexed = list(zip(scores, range(len(to_score))))
indexed.sort(key=lambda x: x[0], reverse=True)
k = 60
combined: list[tuple[float, int]] = []
for idx, (h, rr_score) in enumerate(zip(to_score, scores)):
es_rank = idx + 1
rr_sorted = sorted(scores, reverse=True)
rr_rank = rr_sorted.index(rr_score) + 1
rrf = 1.0 / (k + es_rank) + 1.0 / (k + rr_rank)
combined.append((rrf, idx))
combined.sort(key=lambda x: x[0], reverse=True)
new_order: list = []
for sc, idx in indexed:
for rrf, idx in combined:
h = to_score[idx]
h["_rerank_score"] = float(sc)
h["_rerank_score"] = float(scores[idx])
h["_rrf_score"] = float(rrf)
new_order.append(h)
rest = hits[RERANK_CANDIDATES:]
return new_order + rest, True
except Exception as e:
log.warning("rerank failed: %s", e)
log.warning("rerank failed (fallback to ES): %s", e)
return hits, False
@ -435,7 +448,10 @@ def handle_rag_search(query: str, top_k: int = 8, **kw):
continue
seen_docs.add(dk)
if "_rerank_score" in h:
if "_rrf_score" in h:
score = float(h["_rrf_score"])
score_label = "RRF"
elif "_rerank_score" in h:
score = float(h["_rerank_score"])
score_label = "Rerank"
else: