26794aab09
* 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)
697 lines
28 KiB
Python
697 lines
28 KiB
Python
"""
|
||
Retrieval module for 4-way parallel search.
|
||
|
||
Implements:
|
||
1. Semantic retrieval (vector similarity)
|
||
2. BM25 retrieval (keyword/full-text search)
|
||
3. Graph retrieval (via pluggable GraphRetriever interface)
|
||
4. Temporal retrieval (time-aware search with spreading)
|
||
"""
|
||
|
||
import asyncio
|
||
import logging
|
||
import re
|
||
from dataclasses import dataclass, field
|
||
from datetime import UTC, datetime
|
||
from typing import Optional
|
||
|
||
from ...config import get_config
|
||
from ..db_utils import acquire_with_retry
|
||
from ..memory_engine import fq_table
|
||
from .graph_retrieval import GraphRetriever
|
||
from .link_expansion_retrieval import LinkExpansionRetriever
|
||
from .tags import TagGroup, TagsMatch, build_tag_groups_where_clause, build_tags_where_clause_simple
|
||
from .types import GraphRetrievalTimings, RetrievalResult
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
def tokenize_query(query_text: str) -> list[str]:
|
||
"""Normalize query text and split into BM25 tokens.
|
||
|
||
Strips punctuation, lowercases, and splits on whitespace.
|
||
Returns an empty list when the query contains no word characters.
|
||
"""
|
||
return re.sub(r"[^\w\s]", " ", query_text.lower()).split()
|
||
|
||
|
||
@dataclass
|
||
class ParallelRetrievalResult:
|
||
"""Result from parallel retrieval across all methods."""
|
||
|
||
semantic: list[RetrievalResult]
|
||
bm25: list[RetrievalResult]
|
||
graph: list[RetrievalResult]
|
||
temporal: list[RetrievalResult] | None
|
||
timings: dict[str, float] = field(default_factory=dict)
|
||
temporal_constraint: tuple | None = None # (start_date, end_date)
|
||
graph_timings: list[GraphRetrievalTimings] = field(
|
||
default_factory=list
|
||
) # Graph retrieval sub-step timings per fact type
|
||
max_conn_wait: float = 0.0 # Maximum connection acquisition wait time across all methods
|
||
|
||
|
||
@dataclass
|
||
class MultiFactTypeRetrievalResult:
|
||
"""Result from retrieval across all fact types."""
|
||
|
||
# Results per fact type
|
||
results_by_fact_type: dict[str, ParallelRetrievalResult]
|
||
# Aggregate timings
|
||
timings: dict[str, float] = field(default_factory=dict)
|
||
# Max connection wait across all operations
|
||
max_conn_wait: float = 0.0
|
||
|
||
|
||
# Default graph retriever instance (can be overridden)
|
||
_default_graph_retriever: GraphRetriever | None = None
|
||
|
||
|
||
def get_default_graph_retriever() -> GraphRetriever:
|
||
"""Get or create the default graph retriever based on config."""
|
||
global _default_graph_retriever
|
||
if _default_graph_retriever is None:
|
||
config = get_config()
|
||
retriever_type = config.graph_retriever.lower()
|
||
if retriever_type == "link_expansion":
|
||
_default_graph_retriever = LinkExpansionRetriever()
|
||
logger.info("Using LinkExpansion graph retriever")
|
||
else:
|
||
logger.warning(f"Unknown graph retriever '{retriever_type}', falling back to link_expansion")
|
||
_default_graph_retriever = LinkExpansionRetriever()
|
||
return _default_graph_retriever
|
||
|
||
|
||
def set_default_graph_retriever(retriever: GraphRetriever) -> None:
|
||
"""Set the default graph retriever (for configuration/testing)."""
|
||
global _default_graph_retriever
|
||
_default_graph_retriever = retriever
|
||
|
||
|
||
async def retrieve_semantic_bm25_combined(
|
||
conn,
|
||
query_emb_str: str,
|
||
query_text: str,
|
||
bank_id: str,
|
||
fact_types: list[str],
|
||
limit: int,
|
||
tags: list[str] | None = None,
|
||
tags_match: TagsMatch = "any",
|
||
tag_groups: list[TagGroup] | None = None,
|
||
) -> dict[str, tuple[list[RetrievalResult], list[RetrievalResult]]]:
|
||
"""
|
||
Combined semantic + BM25 retrieval for multiple fact types in a single query.
|
||
|
||
Uses UNION ALL of per-fact_type subqueries so that each arm has its own
|
||
ORDER BY ... LIMIT, enabling the partial HNSW indexes per fact_type instead
|
||
of forcing a full sequential scan (which the previous window-function approach
|
||
caused by using PARTITION BY inside ROW_NUMBER()).
|
||
|
||
Requires partial HNSW indexes per fact_type (idx_mu_emb_world,
|
||
idx_mu_emb_observation, idx_mu_emb_experience), created automatically by
|
||
Alembic migration a3b4c5d6e7f8_add_partial_hnsw_indexes.py.
|
||
|
||
HNSW is approximate — semantic arms over-fetch by 5x (min 100) and trim to
|
||
limit in Python to compensate. ef_search=200 is set globally on pool
|
||
connections at init time (see memory_engine.py) to improve recall on sparse
|
||
graphs.
|
||
|
||
fact_type values are inlined as literals (safe: they come from a controlled
|
||
internal enum, never from user input).
|
||
|
||
Args:
|
||
conn: Database connection
|
||
query_emb_str: Query embedding as string
|
||
query_text: Query text for BM25
|
||
bank_id: Bank ID
|
||
fact_types: List of fact types to retrieve
|
||
limit: Maximum results per method per fact type
|
||
tags: Optional tags to filter by
|
||
tags_match: Tag matching mode
|
||
|
||
Returns:
|
||
Dict mapping fact_type -> (semantic_results, bm25_results)
|
||
"""
|
||
result_dict: dict[str, tuple[list[RetrievalResult], list[RetrievalResult]]] = {ft: ([], []) for ft in fact_types}
|
||
|
||
tokens = tokenize_query(query_text)
|
||
|
||
# Over-fetch for HNSW approximation; semantic results trimmed to limit in Python.
|
||
hnsw_fetch = max(limit * 5, 100)
|
||
|
||
cols = (
|
||
"id, text, context, event_date, occurred_start, occurred_end, mentioned_at, "
|
||
"fact_type, document_id, chunk_id, tags, metadata, proof_count"
|
||
)
|
||
table = fq_table("memory_units")
|
||
|
||
# --- Parameter layout ---
|
||
# $1 = query_emb_str (semantic arms)
|
||
# $2 = bank_id
|
||
# When tokens present:
|
||
# $3 = limit (BM25 LIMIT; semantic uses inlined hnsw_fetch literal)
|
||
# $4 = bm25_text
|
||
# $5 = tags (if present)
|
||
# $6+ = tag_groups params (one per leaf)
|
||
# When no tokens ($3 is skipped — not included in params to avoid type inference gap):
|
||
# $3 = tags (if present)
|
||
# $4+ = tag_groups params (one per leaf)
|
||
tags_param_idx = 5 if tokens else 3
|
||
tags_clause = build_tags_where_clause_simple(tags, tags_param_idx, match=tags_match)
|
||
|
||
# tag_groups params start immediately after the tags param slot
|
||
tag_groups_param_start = tags_param_idx + (1 if tags else 0)
|
||
groups_clause, groups_params, _ = build_tag_groups_where_clause(tag_groups, tag_groups_param_start)
|
||
|
||
# --- Semantic UNION ALL arms (one per fact_type) ---
|
||
# Each arm has its own ORDER BY embedding <=> $1 LIMIT {hnsw_fetch}, which
|
||
# lets the planner use the partial HNSW index for that fact_type.
|
||
sem_arms = []
|
||
for ft in fact_types:
|
||
sem_arms.append(
|
||
f"(SELECT {cols},"
|
||
f" 1 - (embedding <=> $1::vector) AS similarity,"
|
||
f" NULL::float AS bm25_score,"
|
||
f" 'semantic' AS source"
|
||
f" FROM {table}"
|
||
f" WHERE bank_id = $2"
|
||
f" AND fact_type = '{ft}'"
|
||
f" AND embedding IS NOT NULL"
|
||
f" AND (1 - (embedding <=> $1::vector)) >= 0.3"
|
||
f" {tags_clause}"
|
||
f" {groups_clause}"
|
||
f" ORDER BY embedding <=> $1::vector"
|
||
f" LIMIT {hnsw_fetch})"
|
||
)
|
||
|
||
arms = sem_arms
|
||
|
||
# --- BM25 UNION ALL arms (one per fact_type, only when tokens present) ---
|
||
if tokens:
|
||
config = get_config()
|
||
if config.text_search_extension == "vchord":
|
||
bm25_score_expr = (
|
||
"search_vector <&> to_bm25query('idx_memory_units_text_search', tokenize($4, 'llmlingua2'))"
|
||
)
|
||
bm25_order_by = f"{bm25_score_expr} DESC"
|
||
bm25_where_filter = ""
|
||
bm25_text_param: str = query_text
|
||
elif config.text_search_extension == "pg_textsearch":
|
||
bm25_score_expr = "-(text <@> to_bm25query($4, 'idx_memory_units_text_search'))"
|
||
bm25_order_by = "text <@> to_bm25query($4, 'idx_memory_units_text_search') ASC"
|
||
bm25_where_filter = ""
|
||
bm25_text_param = query_text
|
||
else: # native
|
||
query_tsquery = " | ".join(tokens)
|
||
bm25_score_expr = "ts_rank_cd(search_vector, to_tsquery('english', $4))"
|
||
bm25_order_by = f"{bm25_score_expr} DESC"
|
||
bm25_where_filter = "AND search_vector @@ to_tsquery('english', $4)"
|
||
bm25_text_param = query_tsquery
|
||
|
||
for ft in fact_types:
|
||
arms.append(
|
||
f"(SELECT {cols},"
|
||
f" NULL::float AS similarity,"
|
||
f" {bm25_score_expr} AS bm25_score,"
|
||
f" 'bm25' AS source"
|
||
f" FROM {table}"
|
||
f" WHERE bank_id = $2"
|
||
f" AND fact_type = '{ft}'"
|
||
f" {bm25_where_filter}"
|
||
f" {tags_clause}"
|
||
f" {groups_clause}"
|
||
f" ORDER BY {bm25_order_by}"
|
||
f" LIMIT $3)"
|
||
)
|
||
|
||
query = "\nUNION ALL\n".join(arms)
|
||
|
||
params: list = [query_emb_str, bank_id]
|
||
if tokens:
|
||
params.append(limit) # $3: BM25 LIMIT (only referenced when tokens are present)
|
||
params.append(bm25_text_param) # $4
|
||
if tags:
|
||
params.append(tags)
|
||
params.extend(groups_params)
|
||
|
||
rows = await conn.fetch(query, *params)
|
||
|
||
# Group results; trim semantic to limit (over-fetched for HNSW approximation).
|
||
sem_counts: dict[str, int] = {ft: 0 for ft in fact_types}
|
||
for r in rows:
|
||
row = dict(r)
|
||
source = row.pop("source")
|
||
ft = row.get("fact_type")
|
||
if ft not in result_dict:
|
||
continue
|
||
if source == "semantic":
|
||
if sem_counts[ft] < limit:
|
||
result_dict[ft][0].append(RetrievalResult.from_db_row(row))
|
||
sem_counts[ft] += 1
|
||
else:
|
||
result_dict[ft][1].append(RetrievalResult.from_db_row(row))
|
||
|
||
return result_dict
|
||
|
||
|
||
async def retrieve_temporal_combined(
|
||
conn,
|
||
query_emb_str: str,
|
||
bank_id: str,
|
||
fact_types: list[str],
|
||
start_date: datetime,
|
||
end_date: datetime,
|
||
budget: int,
|
||
semantic_threshold: float = 0.1,
|
||
tags: list[str] | None = None,
|
||
tags_match: TagsMatch = "any",
|
||
tag_groups: list[TagGroup] | None = None,
|
||
) -> dict[str, list[RetrievalResult]]:
|
||
"""
|
||
Temporal retrieval for multiple fact types in a single query.
|
||
|
||
Batches the entry point query using window functions to get top-N per fact type,
|
||
then runs spreading for each fact type.
|
||
|
||
Args:
|
||
conn: Database connection
|
||
query_emb_str: Query embedding as string
|
||
bank_id: Bank ID
|
||
fact_types: List of fact types to retrieve
|
||
start_date: Start of time range
|
||
end_date: End of time range
|
||
budget: Node budget for spreading per fact type
|
||
semantic_threshold: Minimum semantic similarity to include
|
||
|
||
Returns:
|
||
Dict mapping fact_type -> list of RetrievalResult
|
||
"""
|
||
from ..memory_engine import fq_table
|
||
|
||
# Ensure dates are timezone-aware
|
||
if start_date.tzinfo is None:
|
||
start_date = start_date.replace(tzinfo=UTC)
|
||
if end_date.tzinfo is None:
|
||
end_date = end_date.replace(tzinfo=UTC)
|
||
|
||
# Build tags clause
|
||
# Entry point query: fixed params are $1-$6, tags at $7
|
||
tags_clause = build_tags_where_clause_simple(tags, 7, match=tags_match)
|
||
tag_groups_param_start = 7 + (1 if tags else 0)
|
||
groups_clause, groups_params, _ = build_tag_groups_where_clause(tag_groups, tag_groups_param_start)
|
||
params: list = [query_emb_str, bank_id, fact_types, start_date, end_date, semantic_threshold]
|
||
if tags:
|
||
params.append(tags)
|
||
params.extend(groups_params)
|
||
|
||
# Two-phase entry point query:
|
||
# Phase 1 (date_ranked): rank by date only — no embedding computation — for all units in
|
||
# the temporal window. This lets the planner use date indexes for filtering.
|
||
# Phase 2 (sim_ranked): join back to memory_units for only the top-50-per-type candidates
|
||
# and compute embedding similarity for that small set (≤ 50 × len(fact_types) rows).
|
||
# This avoids computing embedding distances for potentially thousands of date-range rows.
|
||
entry_points = await conn.fetch(
|
||
f"""
|
||
WITH date_ranked AS MATERIALIZED (
|
||
SELECT id, fact_type,
|
||
ROW_NUMBER() OVER (
|
||
PARTITION BY fact_type
|
||
ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC NULLS LAST
|
||
) AS rn
|
||
FROM {fq_table("memory_units")}
|
||
WHERE bank_id = $2
|
||
AND fact_type = ANY($3)
|
||
AND embedding IS NOT NULL
|
||
AND (
|
||
(occurred_start IS NOT NULL AND occurred_end IS NOT NULL
|
||
AND occurred_start <= $5 AND occurred_end >= $4)
|
||
OR
|
||
(mentioned_at IS NOT NULL AND mentioned_at BETWEEN $4 AND $5)
|
||
OR
|
||
(occurred_start IS NOT NULL AND occurred_start BETWEEN $4 AND $5)
|
||
OR
|
||
(occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
|
||
)
|
||
{tags_clause}
|
||
{groups_clause}
|
||
),
|
||
sim_ranked AS (
|
||
SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.proof_count, mu.document_id, mu.chunk_id, mu.tags, mu.metadata,
|
||
1 - (mu.embedding <=> $1::vector) AS similarity,
|
||
ROW_NUMBER() OVER (PARTITION BY mu.fact_type ORDER BY mu.embedding <=> $1::vector) AS sim_rn
|
||
FROM date_ranked dr
|
||
JOIN {fq_table("memory_units")} mu ON mu.id = dr.id
|
||
WHERE dr.rn <= 50
|
||
AND (1 - (mu.embedding <=> $1::vector)) >= $6
|
||
)
|
||
SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, fact_type, proof_count, document_id, chunk_id, tags, metadata, similarity
|
||
FROM sim_ranked
|
||
WHERE sim_rn <= 10
|
||
""",
|
||
*params,
|
||
)
|
||
|
||
if not entry_points:
|
||
return {ft: [] for ft in fact_types}
|
||
|
||
# Group entry points by fact type
|
||
entries_by_ft: dict[str, list] = {ft: [] for ft in fact_types}
|
||
for ep in entry_points:
|
||
ft = ep["fact_type"]
|
||
if ft in entries_by_ft:
|
||
entries_by_ft[ft].append(ep)
|
||
|
||
# Calculate shared temporal parameters
|
||
total_days = (end_date - start_date).total_seconds() / 86400
|
||
mid_date = start_date + (end_date - start_date) / 2
|
||
|
||
# Process each fact type (spreading needs to stay per fact type due to link filtering)
|
||
results_by_ft: dict[str, list[RetrievalResult]] = {}
|
||
|
||
for ft in fact_types:
|
||
ft_entry_points = entries_by_ft.get(ft, [])
|
||
if not ft_entry_points:
|
||
results_by_ft[ft] = []
|
||
continue
|
||
|
||
results = []
|
||
visited = set()
|
||
node_scores = {}
|
||
|
||
# Process entry points
|
||
for ep in ft_entry_points:
|
||
unit_id = str(ep["id"])
|
||
visited.add(unit_id)
|
||
|
||
# Calculate temporal proximity
|
||
best_date = None
|
||
if ep["occurred_start"] is not None and ep["occurred_end"] is not None:
|
||
best_date = ep["occurred_start"] + (ep["occurred_end"] - ep["occurred_start"]) / 2
|
||
elif ep["occurred_start"] is not None:
|
||
best_date = ep["occurred_start"]
|
||
elif ep["occurred_end"] is not None:
|
||
best_date = ep["occurred_end"]
|
||
elif ep["mentioned_at"] is not None:
|
||
best_date = ep["mentioned_at"]
|
||
|
||
if best_date:
|
||
days_from_mid = abs((best_date - mid_date).total_seconds() / 86400)
|
||
temporal_proximity = 1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
|
||
else:
|
||
temporal_proximity = 0.5
|
||
|
||
ep_result = RetrievalResult.from_db_row(dict(ep))
|
||
ep_result.temporal_score = temporal_proximity
|
||
ep_result.temporal_proximity = temporal_proximity
|
||
results.append(ep_result)
|
||
node_scores[unit_id] = (ep["similarity"], 1.0)
|
||
|
||
# Spreading through temporal links (same as single-fact-type version)
|
||
frontier = list(node_scores.keys())
|
||
budget_remaining = budget - len(ft_entry_points)
|
||
batch_size = 20
|
||
# Per-source neighbor limit: lets the planner use the composite index
|
||
# (from_unit_id, link_type, weight DESC) with early termination, avoiding
|
||
# a full scan of all links from all source nodes before sorting.
|
||
per_source_limit = 10
|
||
# Safety cap on BFS iterations to prevent runaway spreading in dense graphs.
|
||
max_iterations = 5
|
||
iteration = 0
|
||
|
||
# Build tags clause for spreading (use param 7 since 1-6 are used)
|
||
spreading_tags_clause = build_tags_where_clause_simple(tags, 7, table_alias="mu.", match=tags_match)
|
||
spreading_groups_param_start = 7 + (1 if tags else 0)
|
||
spreading_groups_clause, spreading_groups_params, _ = build_tag_groups_where_clause(
|
||
tag_groups, spreading_groups_param_start, table_alias="mu."
|
||
)
|
||
|
||
while frontier and budget_remaining > 0 and iteration < max_iterations:
|
||
iteration += 1
|
||
batch_ids = frontier[:batch_size]
|
||
frontier = frontier[batch_size:]
|
||
|
||
# $1=query_emb, $2=batch_ids, $3=fact_type, $4=threshold, $5=per_source_limit, $6=bank_id, $7=tags, $M+=tag_groups
|
||
spreading_params = [query_emb_str, batch_ids, ft, semantic_threshold, per_source_limit, bank_id]
|
||
if tags:
|
||
spreading_params.append(tags)
|
||
spreading_params.extend(spreading_groups_params)
|
||
|
||
# LATERAL join: for each source node, fetch top-K neighbors by weight using
|
||
# the existing idx_memory_links_from_type_weight index with early-exit semantics.
|
||
# This avoids scanning all temporal links from all source nodes before sorting.
|
||
# bank_id on memory_units lets the planner use idx_memory_units_bank_fact_type.
|
||
neighbors = await conn.fetch(
|
||
f"""
|
||
SELECT src.from_unit_id, mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.fact_type, mu.document_id, mu.chunk_id, mu.tags, mu.metadata,
|
||
l.weight, l.link_type,
|
||
1 - (mu.embedding <=> $1::vector) AS similarity
|
||
FROM unnest($2::uuid[]) AS src(from_unit_id)
|
||
CROSS JOIN LATERAL (
|
||
SELECT ml.to_unit_id, ml.weight, ml.link_type
|
||
FROM {fq_table("memory_links")} ml
|
||
WHERE ml.from_unit_id = src.from_unit_id
|
||
AND ml.link_type IN ('temporal', 'causes', 'caused_by', 'enables', 'prevents')
|
||
AND ml.weight >= 0.1
|
||
ORDER BY ml.weight DESC
|
||
LIMIT $5
|
||
) l
|
||
JOIN {fq_table("memory_units")} mu ON mu.id = l.to_unit_id
|
||
WHERE mu.bank_id = $6
|
||
AND mu.fact_type = $3
|
||
AND mu.embedding IS NOT NULL
|
||
AND (1 - (mu.embedding <=> $1::vector)) >= $4
|
||
{spreading_tags_clause}
|
||
{spreading_groups_clause}
|
||
""",
|
||
*spreading_params,
|
||
)
|
||
|
||
for n in neighbors:
|
||
neighbor_id = str(n["id"])
|
||
if neighbor_id in visited:
|
||
continue
|
||
|
||
visited.add(neighbor_id)
|
||
budget_remaining -= 1
|
||
|
||
parent_id = str(n["from_unit_id"])
|
||
_, parent_temporal_score = node_scores.get(parent_id, (0.5, 0.5))
|
||
|
||
neighbor_best_date = None
|
||
if n["occurred_start"] is not None and n["occurred_end"] is not None:
|
||
neighbor_best_date = n["occurred_start"] + (n["occurred_end"] - n["occurred_start"]) / 2
|
||
elif n["occurred_start"] is not None:
|
||
neighbor_best_date = n["occurred_start"]
|
||
elif n["occurred_end"] is not None:
|
||
neighbor_best_date = n["occurred_end"]
|
||
elif n["mentioned_at"] is not None:
|
||
neighbor_best_date = n["mentioned_at"]
|
||
|
||
if neighbor_best_date:
|
||
days_from_mid = abs((neighbor_best_date - mid_date).total_seconds() / 86400)
|
||
neighbor_temporal_proximity = (
|
||
1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
|
||
)
|
||
else:
|
||
neighbor_temporal_proximity = 0.3
|
||
|
||
link_type = n["link_type"]
|
||
if link_type in ("causes", "caused_by"):
|
||
causal_boost = 2.0
|
||
elif link_type in ("enables", "prevents"):
|
||
causal_boost = 1.5
|
||
else:
|
||
causal_boost = 1.0
|
||
|
||
propagated_temporal = parent_temporal_score * n["weight"] * causal_boost * 0.7
|
||
combined_temporal = max(neighbor_temporal_proximity, propagated_temporal)
|
||
|
||
neighbor_result = RetrievalResult.from_db_row(dict(n))
|
||
neighbor_result.temporal_score = combined_temporal
|
||
neighbor_result.temporal_proximity = neighbor_temporal_proximity
|
||
results.append(neighbor_result)
|
||
|
||
if budget_remaining > 0 and combined_temporal > 0.2:
|
||
node_scores[neighbor_id] = (n["similarity"], combined_temporal)
|
||
frontier.append(neighbor_id)
|
||
|
||
if budget_remaining <= 0:
|
||
break
|
||
|
||
results_by_ft[ft] = results
|
||
|
||
return results_by_ft
|
||
|
||
|
||
async def retrieve_all_fact_types_parallel(
|
||
pool,
|
||
query_text: str,
|
||
query_embedding_str: str,
|
||
bank_id: str,
|
||
fact_types: list[str],
|
||
thinking_budget: int,
|
||
question_date: datetime | None = None,
|
||
query_analyzer: Optional["QueryAnalyzer"] = None,
|
||
graph_retriever: GraphRetriever | None = None,
|
||
tags: list[str] | None = None,
|
||
tags_match: TagsMatch = "any",
|
||
tag_groups: list[TagGroup] | None = None,
|
||
) -> MultiFactTypeRetrievalResult:
|
||
"""
|
||
Optimized retrieval for multiple fact types using batched queries.
|
||
|
||
This reduces database round-trips by:
|
||
1. Combining semantic + BM25 into one CTE query for ALL fact types (1 query instead of 2N)
|
||
2. Running graph retrieval per fact type in parallel (N parallel tasks)
|
||
3. Running temporal retrieval per fact type in parallel (N parallel tasks)
|
||
|
||
Args:
|
||
pool: Database connection pool
|
||
query_text: Query text
|
||
query_embedding_str: Query embedding as string
|
||
bank_id: Bank ID
|
||
fact_types: List of fact types to retrieve
|
||
thinking_budget: Budget for graph traversal and retrieval limits
|
||
question_date: Optional date when question was asked (for temporal filtering)
|
||
query_analyzer: Query analyzer to use (defaults to TransformerQueryAnalyzer)
|
||
graph_retriever: Graph retrieval strategy (defaults to configured retriever)
|
||
|
||
Returns:
|
||
MultiFactTypeRetrievalResult with results organized by fact type
|
||
"""
|
||
import time
|
||
|
||
retriever = graph_retriever or get_default_graph_retriever()
|
||
start_time = time.time()
|
||
timings: dict[str, float] = {}
|
||
|
||
# Step 1: Extract temporal constraint first (CPU work, no DB)
|
||
# Do this before DB queries so we know if we need temporal retrieval
|
||
temporal_extraction_start = time.time()
|
||
from .temporal_extraction import extract_temporal_constraint
|
||
|
||
temporal_constraint = extract_temporal_constraint(query_text, reference_date=question_date, analyzer=query_analyzer)
|
||
temporal_extraction_time = time.time() - temporal_extraction_start
|
||
timings["temporal_extraction"] = temporal_extraction_time
|
||
|
||
# Step 2: Run semantic + BM25 + temporal combined in ONE connection!
|
||
# This reduces connection usage from 2 to 1 for these operations
|
||
semantic_bm25_start = time.time()
|
||
temporal_results_by_ft: dict[str, list[RetrievalResult]] = {}
|
||
temporal_time = 0.0
|
||
|
||
async with acquire_with_retry(pool) as conn:
|
||
conn_wait = time.time() - semantic_bm25_start
|
||
|
||
# Semantic + BM25 combined
|
||
semantic_bm25_results = await retrieve_semantic_bm25_combined(
|
||
conn,
|
||
query_embedding_str,
|
||
query_text,
|
||
bank_id,
|
||
fact_types,
|
||
thinking_budget,
|
||
tags=tags,
|
||
tags_match=tags_match,
|
||
tag_groups=tag_groups,
|
||
)
|
||
semantic_bm25_time = time.time() - semantic_bm25_start
|
||
|
||
# Temporal combined (if constraint detected) - same connection!
|
||
if temporal_constraint:
|
||
tc_start, tc_end = temporal_constraint
|
||
temporal_start = time.time()
|
||
temporal_results_by_ft = await retrieve_temporal_combined(
|
||
conn,
|
||
query_embedding_str,
|
||
bank_id,
|
||
fact_types,
|
||
tc_start,
|
||
tc_end,
|
||
budget=thinking_budget,
|
||
semantic_threshold=0.1,
|
||
tags=tags,
|
||
tags_match=tags_match,
|
||
tag_groups=tag_groups,
|
||
)
|
||
temporal_time = time.time() - temporal_start
|
||
|
||
timings["semantic_bm25_combined"] = semantic_bm25_time
|
||
timings["temporal_combined"] = temporal_time
|
||
|
||
# Step 3: Run graph retrieval for each fact type in parallel
|
||
async def run_graph_for_fact_type(
|
||
ft: str,
|
||
) -> tuple[str, list[RetrievalResult], float, GraphRetrievalTimings | None]:
|
||
graph_start = time.time()
|
||
results, graph_timing = await retriever.retrieve(
|
||
pool=pool,
|
||
query_embedding_str=query_embedding_str,
|
||
bank_id=bank_id,
|
||
fact_type=ft,
|
||
budget=thinking_budget,
|
||
query_text=query_text,
|
||
semantic_seeds=None,
|
||
temporal_seeds=None,
|
||
tags=tags,
|
||
tags_match=tags_match,
|
||
tag_groups=tag_groups,
|
||
)
|
||
return ft, results, time.time() - graph_start, graph_timing
|
||
|
||
# Run graph for all fact types in parallel
|
||
graph_tasks = [run_graph_for_fact_type(ft) for ft in fact_types]
|
||
graph_results_list = await asyncio.gather(*graph_tasks)
|
||
|
||
# Organize results by fact type
|
||
results_by_fact_type: dict[str, ParallelRetrievalResult] = {}
|
||
max_conn_wait = conn_wait # Single connection for semantic+bm25+temporal
|
||
all_graph_timings: list[GraphRetrievalTimings] = []
|
||
|
||
for ft in fact_types:
|
||
# Get semantic + bm25 results for this fact type
|
||
semantic_results, bm25_results = semantic_bm25_results.get(ft, ([], []))
|
||
|
||
# Find graph results for this fact type
|
||
graph_results = []
|
||
graph_time = 0.0
|
||
graph_timing = None
|
||
for gr in graph_results_list:
|
||
if gr[0] == ft:
|
||
graph_results = gr[1]
|
||
graph_time = gr[2]
|
||
graph_timing = gr[3]
|
||
if graph_timing:
|
||
all_graph_timings.append(graph_timing)
|
||
break
|
||
|
||
# Get temporal results for this fact type from combined result
|
||
temporal_results = temporal_results_by_ft.get(ft) if temporal_constraint else None
|
||
if temporal_results is not None and len(temporal_results) == 0:
|
||
temporal_results = None
|
||
|
||
results_by_fact_type[ft] = ParallelRetrievalResult(
|
||
semantic=semantic_results,
|
||
bm25=bm25_results,
|
||
graph=graph_results,
|
||
temporal=temporal_results,
|
||
timings={
|
||
"semantic": semantic_bm25_time / 2, # Approximate split
|
||
"bm25": semantic_bm25_time / 2,
|
||
"graph": graph_time,
|
||
"temporal": temporal_time, # Same for all fact types (single query)
|
||
"temporal_extraction": temporal_extraction_time,
|
||
},
|
||
temporal_constraint=temporal_constraint,
|
||
graph_timings=[graph_timing] if graph_timing else [],
|
||
max_conn_wait=max_conn_wait,
|
||
)
|
||
|
||
total_time = time.time() - start_time
|
||
timings["total"] = total_time
|
||
|
||
return MultiFactTypeRetrievalResult(
|
||
results_by_fact_type=results_by_fact_type,
|
||
timings=timings,
|
||
max_conn_wait=max_conn_wait,
|
||
)
|