jason/mempalace
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(searcher): add candidate_strategy="union" for vector∪BM25 reranking pool

Browse Source 
Default search behavior is unchanged. Opt-in candidate_strategy="union"
also pulls top-K BM25-only candidates from sqlite FTS5 and merges them
into the rerank pool, catching docs with strong BM25 signal that the
vector index didn't surface in the over-fetch window.

Motivation
----------
The current hybrid path gathers candidates from the vector index only
(n_results * 3 over-fetch), then BM25-reranks within them. When the
query embeds close to the wrong content semantically, the right doc
never enters the rerank pool — *no matter how wide the over-fetch*.
Tested on a ~6K-document mixed corpus (knowledge prose + short structured
records): at *30x* over-fetch (~5% of the corpus) the target doc still
didn't surface for narrative-shaped queries targeting terminology guides.
Wider over-fetch isn't the answer; widening the pool's *source* is.

Concrete failure mode: a narrative-shaped query embeds close to records
sharing the same operational vocabulary (other narrative entries in the
corpus). A terminology / style guide is BM25-strong for the query
(rare keywords the guide repeats) but vector-distant. Vector-only
candidates don't include it; BM25 never gets to rerank it. The hybrid
path produces 0.00 recall on a probe that pure BM25 alone scores 1.00 —
the hybrid is worse than its component on the same input.

Behavior change
---------------
* New parameter ``candidate_strategy: str = "vector"`` on ``search_memories``.
  - ``"vector"`` (default): historical behavior, no change.
  - ``"union"``: also fetch top ``n_results * 3`` candidates via the
    existing ``_bm25_only_via_sqlite`` helper, dedupe by source_file,
    merge into the rerank pool. BM25-only candidates carry
    ``distance=None`` so they're scored on BM25 contribution alone
    (vec_sim coerces to 0).
* ``_hybrid_rank`` now handles ``distance=None`` explicitly, scoring
  such candidates as vector-unknown (vec_sim=0) rather than treating
  it as max-distance via shim.
* New strategies register via ``_CANDIDATE_MERGERS``; dispatch is in
  ``_apply_candidate_strategy`` so ``search_memories`` stays under the
  C901 complexity ceiling.

Bench numbers (~6K-doc internal mixed corpus, recall@10, 5 probes spanning
policy-exception lookup, temporal-decay, style retrieval, set-difference,
and pattern-recognition):

                              baseline ("vector")   "union"
  policy-exception probe        0.00                  0.50    +0.50
  temporal-decay probe          0.17                  0.50    +0.33
  style-retrieval probe         0.00                  1.00    +1.00 (PASSES)
  set-difference probe          0.00–0.06             0.06–0.09  ~
  pattern-recog probe           0.64 (stable)         0.50–0.71  variance, typ. +0.07
  macro recall                  0.16–0.17             0.51–0.56  +0.34 to +0.40

The pattern-recog variance points at a related issue worth a separate PR:
``_hybrid_rank`` computes BM25 IDF over the candidate set. Adding new
candidates re-normalizes BM25 for *existing* candidates non-monotonically.
Stable corpus-wide BM25 would remove this. Out of scope here.

Tests
-----
``tests/test_hybrid_candidate_union.py`` (6 tests, all pass):
- default behavior unchanged (explicit ``"vector"`` matches default)
- ``"union"`` surfaces a BM25-strong vector-distant doc
- ``"union"`` doesn't drop docs ``"vector"`` would have found
- empty-palace handling
- invalid ``candidate_strategy`` raises
- ``_hybrid_rank`` tolerates ``distance=None``

Existing ``test_hybrid_search.py`` (5) and ``test_searcher.py`` (27) pass.

Performance note
----------------
Each ``"union"`` query adds one sqlite open + FTS5 MATCH + metadata
fetch (via the existing ``_bm25_only_via_sqlite`` helper, which already
runs as the ``vector_disabled`` fallback path so the code is
well-trodden). Per-query overhead is small but unmeasured at corpus
scale. Default stays ``"vector"`` until a maintainer characterizes the
cost.
This commit is contained in:
Igor Lins e Silva
2026-05-02 00:47:24 -03:00
parent 5ddaf7abf6
commit 6509071b8e
2 changed files with 236 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
mempalace/searcher.py
+103 -1
View File
@@ -134,6 +134,11 @@ def _hybrid_rank(
themselves. Since the absolute scale is unbounded, BM25 is min-max
normalized within the candidate set so weights are commensurable.
Candidates with ``distance=None`` are treated as vector-unknown
(no vector signal available) and scored on BM25 contribution alone.
Used by candidate-union mode to merge BM25-only candidates that the
vector index didn't surface.
Mutates each result dict to add ``bm25_score`` and reorders the list
in place. Returns the same list for convenience.
"""
@@ -147,7 +152,11 @@ def _hybrid_rank(
scored = []
for r, raw, norm in zip(results, bm25_raw, bm25_norm):
vec_sim = max(0.0, 1.0 - r.get("distance", 1.0))
distance = r.get("distance")
if distance is None:
vec_sim = 0.0
else:
vec_sim = max(0.0, 1.0 - distance)
r["bm25_score"] = round(raw, 3)
scored.append((vector_weight * vec_sim + bm25_weight * norm, r))
@@ -545,6 +554,79 @@ def _bm25_only_via_sqlite(
}
def _merge_bm25_union_candidates(
hits: list,
query: str,
palace_path: str,
wing: str,
room: str,
n_results: int,
) -> None:
"""Append top-K BM25-only candidates from sqlite into ``hits`` in place.
Used by ``search_memories(..., candidate_strategy="union")`` to widen
the rerank pool's *source* (not just its size) — vector-only candidate
selection skips docs whose embeddings are far from the query even when
BM25 signal is strong. We dedupe against existing hits by ``source_file``
so vector-side entries (which carry real distance values) win on
collisions; BM25-only additions are marked with ``distance=None`` so
``_hybrid_rank`` scores them on BM25 contribution alone.
"""
try:
bm25_extra = _bm25_only_via_sqlite(
query,
palace_path,
wing=wing,
room=room,
n_results=n_results * 3,
).get("results", [])
except Exception:
logger.debug("candidate_strategy=union: BM25 fetch failed", exc_info=True)
return
seen_sources = {h.get("source_file") for h in hits}
for bh in bm25_extra:
key = bh.get("source_file")
if not key or key == "?" or key in seen_sources:
continue
bh["distance"] = None
bh["effective_distance"] = None
bh["closet_boost"] = 0.0
hits.append(bh)
seen_sources.add(key)
# Strategy dispatch — keeps search_memories' branch count under the
# project's complexity ceiling (C901 max-complexity=25). New strategies
# register here.
_CANDIDATE_MERGERS = {
"vector": None, # default no-op
"union": _merge_bm25_union_candidates,
}
def _apply_candidate_strategy(
strategy: str,
hits: list,
query: str,
palace_path: str,
wing: str,
room: str,
n_results: int,
) -> None:
"""Dispatch to the registered merger for ``strategy``.
Raises ``ValueError`` for unknown strategies. ``"vector"`` is a no-op.
"""
if strategy not in _CANDIDATE_MERGERS:
raise ValueError(
f"candidate_strategy must be one of {tuple(_CANDIDATE_MERGERS)}, " f"got {strategy!r}"
)
merger = _CANDIDATE_MERGERS[strategy]
if merger is not None:
merger(hits, query, palace_path, wing, room, n_results)
def search_memories(
query: str,
palace_path: str,
@@ -553,6 +635,7 @@ def search_memories(
n_results: int = 5,
max_distance: float = 0.0,
vector_disabled: bool = False,
candidate_strategy: str = "vector",
) -> dict:
"""Programmatic search — returns a dict instead of printing.
@@ -572,6 +655,20 @@ def search_memories(
(#1222). Set by the MCP server when the HNSW capacity probe
detects a divergence that would segfault chromadb on segment
load.
candidate_strategy: How candidates for the hybrid re-rank are gathered.
* ``"vector"`` (default) — preserves historical behavior: top
``n_results * 3`` rows from the vector index are the rerank pool.
Cheap; works well when query and target docs agree in the
embedding space.
* ``"union"`` — also pull top ``n_results * 3`` BM25 candidates
from the sqlite FTS5 index and merge them into the rerank pool
(deduped by source_file). Catches docs with strong BM25 signal
that are vector-distant from the query (e.g. terminology guides
looked up by narrative-shaped queries; policy clauses surfaced
by scenario descriptions). Adds one sqlite open + FTS5 MATCH
per query; perf cost is small but unmeasured at corpus scale.
Opt in until the cost is characterized.
"""
if vector_disabled:
return _bm25_only_via_sqlite(
@@ -748,6 +845,11 @@ def search_memories(
h["drawer_index"] = best_idx
h["total_drawers"] = len(ordered_docs)
# Candidate strategy hook: optionally widen the rerank pool's *source*
# before ranking. Default ("vector") is a no-op; "union" merges top-K
# BM25 candidates from sqlite. See `_apply_candidate_strategy`.
_apply_candidate_strategy(candidate_strategy, hits, query, palace_path, wing, room, n_results)
# BM25 hybrid re-rank within the final candidate set.
hits = _hybrid_rank(hits, query)
for h in hits:
tests/test_hybrid_candidate_union.py
+133
View File
@@ -0,0 +1,133 @@
"""Tests for ``candidate_strategy="union"`` in ``search_memories``.
The default ``"vector"`` strategy gathers candidates from the vector index
only. Docs with strong BM25 signal but vector embeddings far from the query
get skipped — terminology guides looked up by narrative-shaped queries are
the canonical case.
The ``"union"`` strategy also pulls top-K BM25-only candidates from sqlite
FTS5 and merges them into the rerank pool. Both signal sources contribute
candidates; the hybrid rerank picks the best from a richer pool.
Default behavior is unchanged ("vector") — these tests exercise opt-in
"union" mode.
"""
from mempalace.palace import get_collection
from mempalace.searcher import search_memories
def _seed_drawers(palace_path):
"""Seed a corpus where the right doc for one query is BM25-strong but
vector-distant.
D1-D3 are short narrative tickets that semantically cluster around
"customer support / order / shipped" vocabulary. D4 is a meta-document
of bullet rules ("brand voice") that contains rare keywords like
"Absolutely" and "apologize" the query repeats verbatim — strong BM25
signal but stylistically far from the narrative tickets.
"""
col = get_collection(palace_path, create=True)
col.upsert(
ids=["D1", "D2", "D3", "D4"],
documents=[
"Customer wrote in asking why their order shipped without "
"the promo sticker. Standard reply explaining the threshold.",
"Order delivery delayed three days; customer requested a "
"refund. Support agent processed return via ticket queue.",
"Customer asked about the missing freebie; the reply "
"explained the campaign mechanics and shipped status.",
"Brand voice rules: dry, sturdy, never effusive. "
"Never 'Absolutely!' Never apologize for policy — explain it. "
"Avoid premium / curated / elevated vocabulary.",
],
metadatas=[
{"wing": "shop", "room": "support", "source_file": "ticket_D1.md"},
{"wing": "shop", "room": "support", "source_file": "ticket_D2.md"},
{"wing": "shop", "room": "support", "source_file": "ticket_D3.md"},
{"wing": "shop", "room": "guides", "source_file": "brand_voice_D4.md"},
],
)
_NARRATIVE_QUERY = (
"A support agent is drafting a reply to a customer asking why their "
"order shipped without a free sticker. Draft the reply, but never say "
"'Absolutely!' and do not apologize for policy."
)
class TestCandidateUnion:
def test_default_vector_strategy_unchanged(self, tmp_path):
"""Default behavior must be identical to omitting the parameter."""
palace = str(tmp_path / "palace")
_seed_drawers(palace)
without = search_memories(_NARRATIVE_QUERY, palace, n_results=5)
with_default = search_memories(
_NARRATIVE_QUERY, palace, n_results=5, candidate_strategy="vector"
)
ids_a = [h["source_file"] for h in without["results"]]
ids_b = [h["source_file"] for h in with_default["results"]]
assert ids_a == ids_b, "explicit candidate_strategy='vector' must match default"
def test_union_surfaces_bm25_strong_vector_distant_doc(self, tmp_path):
"""The brand-voice doc has strong BM25 signal for the query but is
stylistically far from the narrative tickets. Union mode must
retrieve it; vector-only mode is allowed to miss it."""
palace = str(tmp_path / "palace")
_seed_drawers(palace)
result = search_memories(_NARRATIVE_QUERY, palace, n_results=5, candidate_strategy="union")
ids = [h["source_file"] for h in result["results"]]
assert "brand_voice_D4.md" in ids, (
"union mode must surface BM25-strong docs even when vector signal "
f"is weak; got {ids}"
)
def test_union_preserves_vector_hits(self, tmp_path):
"""Union mode must not drop docs that vector-only mode finds —
the rerank pool grows, it doesn't shrink."""
palace = str(tmp_path / "palace")
_seed_drawers(palace)
vector = search_memories(_NARRATIVE_QUERY, palace, n_results=5, candidate_strategy="vector")
union = search_memories(_NARRATIVE_QUERY, palace, n_results=5, candidate_strategy="union")
vec_ids = {h["source_file"] for h in vector["results"]}
union_ids = {h["source_file"] for h in union["results"]}
# In a 4-doc corpus with n_results=5, both should return all 4.
# The invariant is: union should not lose anything vector found.
missing = vec_ids - union_ids
assert not missing, f"union dropped docs that vector found: {missing}"
def test_union_handles_empty_palace(self, tmp_path):
"""No drawers — union mode should return empty results, not crash."""
palace = str(tmp_path / "palace")
get_collection(palace, create=True) # create empty collection
result = search_memories("anything", palace, n_results=5, candidate_strategy="union")
assert result.get("results", []) == []
def test_invalid_candidate_strategy_raises(self, tmp_path):
"""Bad arg should raise rather than silently fall back."""
palace = str(tmp_path / "palace")
_seed_drawers(palace)
import pytest
with pytest.raises(ValueError, match="candidate_strategy"):
search_memories("anything", palace, n_results=5, candidate_strategy="bogus")
class TestHybridRankTolerantOfMissingDistance:
"""``_hybrid_rank`` accepts ``distance=None`` — required for BM25-only
candidates injected by union mode."""
def test_distance_none_scored_as_zero_vector_sim(self):
from mempalace.searcher import _hybrid_rank
results = [
{"text": "alpha beta gamma", "distance": 0.2}, # close vector match
{"text": "alpha alpha alpha", "distance": None}, # BM25-only — heavy term repetition
]
# Query matches "alpha" heavily; the BM25-only candidate with no
# vector signal should still rank competitively on BM25 alone.
ranked = _hybrid_rank(results, "alpha")
assert all("bm25_score" in r for r in ranked), "rerank should add bm25_score"
# Both must survive — neither should crash on distance=None.
assert len(ranked) == 2