fleet-memory/hindsight-api-slim/hindsight_api/engine/search/reranking.py
Abdulkadirklc 26794aab09
feat(recall): add proof_count boost to combined scoring (#821)
* feat(recall): add proof_count boost to combined scoring

Observations with more supporting evidence now rank slightly higher
in recall results. proof_count is threaded through the retrieval
pipeline and applied as a multiplicative boost in reranking:

- types.py: add proof_count field to RetrievalResult
- retrieval.py: include proof_count in SELECT columns
- reranking.py: add log1p-normalized proof_count boost (alpha=0.1)

The boost uses the same multiplicative pattern as recency and temporal
signals. proof_count=1 is neutral, proof_count=50 gives ~+5% boost.
Non-observation fact types are unaffected (neutral 0.5).

* fix(retrieval): Apply proof_count boost to graph and temporal retrieval, normalize scaling

* fix(retrieval): correct proof_norm math to zero-center at count 1

* fix(retrieval): Apply proof_count boost to link_expansion retrieval

* fix: remove BFS zombie, clamp proof_norm to [0,1], fix test comment (log1p->math.log)
2026-04-07 09:32:44 +02:00

212 lines
8.4 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

"""
Cross-encoder neural reranking for search results.
"""
import math
from datetime import datetime, timezone
from .types import MergedCandidate, ScoredResult
UTC = timezone.utc
# Multiplicative boost alphas for recency and temporal proximity.
# Each signal contributes at most ±(alpha/2) relative adjustment to the base CE score,
# so the max combined boost is (1 + alpha/2)^2 ≈ +21% and min is (1 - alpha/2)^2 ≈ -19%.
_RECENCY_ALPHA: float = 0.2
_TEMPORAL_ALPHA: float = 0.2
_PROOF_COUNT_ALPHA: float = 0.1 # Conservative: max ±5% for evidence strength
def apply_combined_scoring(
scored_results: list[ScoredResult],
now: datetime,
recency_alpha: float = _RECENCY_ALPHA,
temporal_alpha: float = _TEMPORAL_ALPHA,
proof_count_alpha: float = _PROOF_COUNT_ALPHA,
) -> None:
"""Apply combined scoring to a list of ScoredResults in-place.
Uses the cross-encoder score as the primary relevance signal, with recency,
temporal proximity, and proof count applied as multiplicative boosts. This
ensures the influence of these secondary signals is always proportional to
the base relevance score, regardless of the cross-encoder model's score
calibration.
Formula::
recency_boost = 1 + recency_alpha * (recency - 0.5) # in [1-α/2, 1+α/2]
temporal_boost = 1 + temporal_alpha * (temporal - 0.5) # in [1-α/2, 1+α/2]
proof_count_boost = 1 + proof_count_alpha * (proof_norm - 0.5) # in [1-α/2, 1+α/2]
combined_score = CE_normalized * recency_boost * temporal_boost * proof_count_boost
proof_norm maps proof_count using a smooth logarithmic curve centered at 0.5,
clamped to [0, 1]:
proof_count=1 → 0.5 + 0 = 0.5 (neutral multiplier)
proof_count=150 → clamped to 1.0 (max +5% boost)
Temporal proximity is treated as neutral (0.5) when not set by temporal retrieval,
so temporal_boost collapses to 1.0 for non-temporal queries.
Proof count is treated as neutral (0.5) when not available (non-observation facts),
so proof_count_boost collapses to 1.0 for world/experience/opinion facts.
Args:
scored_results: Results from the cross-encoder reranker. Mutated in place.
now: Current UTC datetime for recency calculation.
recency_alpha: Max relative recency adjustment (default 0.2 → ±10%).
temporal_alpha: Max relative temporal adjustment (default 0.2 → ±10%).
proof_count_alpha: Max relative proof count adjustment (default 0.1 → ±5%).
"""
if now.tzinfo is None:
now = now.replace(tzinfo=UTC)
for sr in scored_results:
# Recency: linear decay over 365 days → [0.1, 1.0]; neutral 0.5 if no date.
sr.recency = 0.5
if sr.retrieval.occurred_start:
occurred = sr.retrieval.occurred_start
if occurred.tzinfo is None:
occurred = occurred.replace(tzinfo=UTC)
days_ago = (now - occurred).total_seconds() / 86400
sr.recency = max(0.1, min(1.0, 1.0 - (days_ago / 365)))
# Temporal proximity: meaningful only for temporal queries; neutral otherwise.
sr.temporal = sr.retrieval.temporal_proximity if sr.retrieval.temporal_proximity is not None else 0.5
# Proof count: log-normalized evidence strength; neutral for non-observations.
proof_count = sr.retrieval.proof_count
if proof_count is not None and proof_count >= 1:
# Clamp to [0, 1] so extreme counts stay within documented ±5% range
proof_norm = min(1.0, max(0.0, 0.5 + (math.log(proof_count) / 10.0)))
else:
# Neutral baseline is precisely 0.5, ensuring neutral multiplier (1.0)
proof_norm = 0.5
# RRF: kept at 0.0 for trace continuity but excluded from scoring.
# RRF is batch-relative (min-max normalised) and redundant after reranking.
sr.rrf_normalized = 0.0
recency_boost = 1.0 + recency_alpha * (sr.recency - 0.5)
temporal_boost = 1.0 + temporal_alpha * (sr.temporal - 0.5)
proof_count_boost = 1.0 + proof_count_alpha * (proof_norm - 0.5)
sr.combined_score = sr.cross_encoder_score_normalized * recency_boost * temporal_boost * proof_count_boost
sr.weight = sr.combined_score
class CrossEncoderReranker:
"""
Neural reranking using a cross-encoder model.
Configured via environment variables (see cross_encoder.py).
Default local model is cross-encoder/ms-marco-MiniLM-L-6-v2.
"""
def __init__(self, cross_encoder=None):
"""
Initialize cross-encoder reranker.
Args:
cross_encoder: CrossEncoderModel instance. If None, creates one from
environment variables (defaults to local provider)
"""
if cross_encoder is None:
from hindsight_api.engine.cross_encoder import create_cross_encoder_from_env
cross_encoder = create_cross_encoder_from_env()
self.cross_encoder = cross_encoder
self._initialized = False
async def ensure_initialized(self):
"""Ensure the cross-encoder model is initialized (for lazy initialization)."""
if self._initialized:
return
import asyncio
cross_encoder = self.cross_encoder
# For local providers, run in thread pool to avoid blocking event loop
if cross_encoder.provider_name == "local":
loop = asyncio.get_event_loop()
await loop.run_in_executor(None, lambda: asyncio.run(cross_encoder.initialize()))
else:
await cross_encoder.initialize()
self._initialized = True
async def rerank(self, query: str, candidates: list[MergedCandidate]) -> list[ScoredResult]:
"""
Rerank candidates using cross-encoder scores.
Args:
query: Search query
candidates: Merged candidates from RRF
Returns:
List of ScoredResult objects sorted by cross-encoder score
"""
if not candidates:
return []
# Prepare query-document pairs with date information
pairs = []
for candidate in candidates:
retrieval = candidate.retrieval
# Use text + context for better ranking
doc_text = retrieval.text
if retrieval.context:
doc_text = f"{retrieval.context}: {doc_text}"
# Add formatted date information for temporal awareness
if retrieval.occurred_start:
occurred_start = retrieval.occurred_start
# Format in two styles for better model understanding
# 1. ISO format: YYYY-MM-DD
date_iso = occurred_start.strftime("%Y-%m-%d")
# 2. Human-readable: "June 5, 2022"
date_readable = occurred_start.strftime("%B %d, %Y")
# Prepend date to document text
doc_text = f"[Date: {date_readable} ({date_iso})] {doc_text}"
pairs.append([query, doc_text])
# Get cross-encoder scores
scores = await self.cross_encoder.predict(pairs)
# Normalize scores using sigmoid to [0, 1] range
# Cross-encoder returns logits which can be negative
import math
import numpy as np
def sigmoid(x):
return 1 / (1 + np.exp(-x))
normalized_scores = [sigmoid(score) for score in scores]
# Create ScoredResult objects with cross-encoder scores
scored_results = []
for candidate, raw_score, norm_score in zip(candidates, scores, normalized_scores):
# Sanitize NaN scores (cross-encoder can return NaN for certain inputs).
# NaN propagates through all downstream scoring and Pydantic serializes
# NaN as JSON null, which breaks clients expecting numeric values.
raw = float(raw_score)
norm = float(norm_score)
if math.isnan(raw):
raw = 0.0
if math.isnan(norm):
norm = 0.0
scored_result = ScoredResult(
candidate=candidate,
cross_encoder_score=raw,
cross_encoder_score_normalized=norm,
weight=norm, # Initial weight is just cross-encoder score
)
scored_results.append(scored_result)
# Sort by cross-encoder score
scored_results.sort(key=lambda x: x.weight, reverse=True)
return scored_results