Spaces:
Configuration error
Configuration error
File size: 9,245 Bytes
6733714 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 | """
Retrieval Memory (Priority 9)
================================
Logs every query + retrieval outcome to an append-only JSONL file.
Provides a mining function that surfaces "if query mentions X, always retrieve Y" patterns
so they can be fed back into the authority registry.
Two components:
1. RetrievalLogger β appends one JSON line per query (zero latency impact;
writes are fire-and-forget via a background thread).
2. mine_retrieval_patterns() β offline analysis script that reads the log
and produces authority_registry ADDITIONS for topics that are consistently
missing an authority that always appears when manually inspected.
Usage in retriever.py:
from app.retrieval.retrieval_memory import RetrievalLogger
_mem_logger = RetrievalLogger()
# After coverage check, inside search() or supplement_and_rerank():
_mem_logger.log(
query = query,
topic = _taxonomy.get("topics", []),
retrieved = [c.get("rel_path", "") for c in final_chunks[:15]],
coverage_pct = _coverage_result.get("coverage_pct", 0),
missing = _coverage_result.get("missing", []),
)
"""
from __future__ import annotations
import json
import logging
import os
import threading
from collections import defaultdict
from datetime import datetime, timezone
from pathlib import Path
from typing import Any
logger = logging.getLogger(__name__)
# ββ Log file location βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
_DEFAULT_LOG = Path(__file__).resolve().parent.parent.parent / "data" / "retrieval_memory.jsonl"
class RetrievalLogger:
"""
Append-only JSONL logger for retrieval outcomes.
All writes happen on a daemon worker thread β zero latency impact on the
main retrieval path.
Log entry schema (one JSON object per line):
{
"ts": "ISO-8601 UTC timestamp",
"query": "raw query string",
"topics": ["cross_charge", ...], # taxonomy topic matches
"retrieved": ["circulars/199.../...", ...], # top-15 rel_paths
"coverage_pct": 85, # mandatory coverage %
"missing": ["CGST_SEC_25", ...], # authorities NOT found
"duration_ms": 142 # retrieval wall time (optional)
}
"""
def __init__(self, log_path: Path | str | None = None):
self._path = Path(log_path) if log_path else _DEFAULT_LOG
self._path.parent.mkdir(parents=True, exist_ok=True)
self._lock = threading.Lock()
self._queue = []
self._worker = threading.Thread(target=self._flush_loop, daemon=True)
self._worker.start()
logger.info(f"RetrievalLogger: logging to {self._path}")
def log(
self,
query: str,
topics: list[str],
retrieved: list[str],
coverage_pct: int,
missing: list[str],
duration_ms: int = 0,
) -> None:
"""Enqueues a log entry (non-blocking)."""
entry = {
"ts": datetime.now(timezone.utc).isoformat(),
"query": query,
"topics": topics,
"retrieved": retrieved[:20], # cap to avoid bloat
"coverage_pct": coverage_pct,
"missing": missing,
"duration_ms": duration_ms,
}
with self._lock:
self._queue.append(entry)
def _flush_loop(self) -> None:
"""Worker: drains the queue every 5 seconds."""
import time
while True:
time.sleep(5)
self._drain()
def _drain(self) -> None:
with self._lock:
batch, self._queue = self._queue, []
if not batch:
return
try:
with open(self._path, "a", encoding="utf-8") as fh:
for entry in batch:
fh.write(json.dumps(entry, ensure_ascii=False) + "\n")
except Exception as exc:
logger.warning(f"RetrievalLogger flush failed: {exc}")
# ββ Offline pattern miner βββββββββββββββββββββββββββββββββββββββββββββββββββββ
def mine_retrieval_patterns(
log_path: Path | str | None = None,
min_occurrences: int = 5,
min_miss_rate: float = 0.3,
output_path: Path | str | None = None,
) -> dict[str, Any]:
"""
Analyses the retrieval log to find patterns like:
"For topic X, authority Y is missing in 40% of queries β add Y to registry"
Args:
log_path: Path to retrieval_memory.jsonl (default: data/retrieval_memory.jsonl)
min_occurrences: Minimum times a topic must appear to be included in report
min_miss_rate: Fraction of queries where an authority is missing to flag it
output_path: If set, writes JSON report to this path
Returns:
{
"topic_stats": {
"cross_charge": {
"count": 42,
"avg_coverage_pct": 67,
"frequent_missing": [
{"authority": "CGST_SEC_25", "miss_rate": 0.52, "count": 22}
]
}
},
"registry_additions": {
"cross_charge": {
"sections": ["CGST_SEC_25"], # should add to registry
"circulars": ["CIRCULAR_199"],
}
}
}
After running this, review the `registry_additions` section and apply them
to data/authority_registry.json. The legal team can also run this manually
after observing patterns in production.
"""
_log_path = Path(log_path) if log_path else _DEFAULT_LOG
if not _log_path.exists():
return {"error": f"Log file not found: {_log_path}", "topic_stats": {}, "registry_additions": {}}
# Load log
entries = []
with open(_log_path, "r", encoding="utf-8") as fh:
for line in fh:
line = line.strip()
if line:
try:
entries.append(json.loads(line))
except json.JSONDecodeError:
continue
if not entries:
return {"error": "Log is empty", "topic_stats": {}, "registry_additions": {}}
# Aggregate per topic
topic_counts: dict[str, int] = defaultdict(int)
topic_cov_sum: dict[str, int] = defaultdict(int)
topic_miss: dict[str, dict[str, int]] = defaultdict(lambda: defaultdict(int))
for entry in entries:
for topic in entry.get("topics", ["unknown"]):
topic_counts[topic] += 1
topic_cov_sum[topic] += entry.get("coverage_pct", 100)
for miss in entry.get("missing", []):
topic_miss[topic][miss] += 1
# Build stats
topic_stats: dict[str, Any] = {}
registry_additions: dict[str, Any] = {}
for topic, count in sorted(topic_counts.items(), key=lambda x: -x[1]):
if count < min_occurrences:
continue
avg_cov = round(topic_cov_sum[topic] / count)
miss_items = topic_miss.get(topic, {})
frequent_missing = [
{
"authority": auth,
"miss_rate": round(n / count, 2),
"count": n,
}
for auth, n in sorted(miss_items.items(), key=lambda x: -x[1])
if n / count >= min_miss_rate
]
topic_stats[topic] = {
"count": count,
"avg_coverage_pct": avg_cov,
"frequent_missing": frequent_missing,
}
if frequent_missing:
additions: dict[str, list[str]] = {"sections": [], "rules": [], "circulars": []}
for item in frequent_missing:
auth = item["authority"]
if "_SEC_" in auth:
additions["sections"].append(auth)
elif "_RUL_" in auth:
additions["rules"].append(auth)
elif auth.startswith("CIRCULAR_"):
additions["circulars"].append(auth)
if any(additions.values()):
registry_additions[topic] = additions
result = {
"analysed": len(entries),
"first_entry": entries[0]["ts"] if entries else None,
"last_entry": entries[-1]["ts"] if entries else None,
"topic_stats": topic_stats,
"registry_additions": registry_additions,
}
if output_path:
with open(Path(output_path), "w", encoding="utf-8") as fh:
json.dump(result, fh, indent=2, ensure_ascii=False)
logger.info(f"Pattern mining report written to {output_path}")
return result
if __name__ == "__main__":
"""
Quick CLI:
python -m app.retrieval.retrieval_memory
"""
import sys
_out = sys.argv[1] if len(sys.argv) > 1 else None
report = mine_retrieval_patterns(output_path=_out)
print(json.dumps(report, indent=2, ensure_ascii=False))
|