Phase 7+8: eval harness + hybrid retrieval

## Phase 7 — Eval harness

eval/retrievers.py + rag/retrieval.py: Retriever protocol with
DenseRetriever, BM25Retriever, HybridRetriever (RRF k=60),
RerankedRetriever (llama.cpp /v1/rerank). retrievers.py is now a
thin shim re-exporting from rag.retrieval so the MCP server can
use the same code at request time without making eval/ a runtime
dep.

eval/run_eval.py: drives N retrievers against eval/queries.jsonl,
computes MRR / Recall@K / nDCG@K, emits a markdown report with a
summary table + per-query breakdown for the first retriever. Each
query carries expected (source, source_key) tuples — matches the
labels-domain page-level keying.

eval/queries.jsonl: 35 curated queries — 25 brand-name (Warrant,
Huskie, Roundup Custom, Liberty, Authority, Headline, Trivapro,
Poncho, Lorsban, Sencor, Acuron, ...) + 10 intent/semantic
("what controls horseweed before soybean", "fungicide for fusarium
head blight", "rainfast interval for glyphosate", ...).

## Phase 8 — Hybrid retrieval (BM25 + dense + RRF)

docs_mcp/server.py: search_docs now branches on HYBRID_SEARCH env.
When on, _search_chunks runs both Chroma + BM25 (rag/bm25.py
existing impl), fuses on chunk_id with reciprocal-rank-fusion
(RRF k=60), and returns the combined pool. Dense-only path
unchanged when HYBRID_SEARCH is unset. The rendering layer
(_format_hit) is untouched.

The RERANK_URL hook is also wired (_rerank_pool sends docs to
llama.cpp /v1/rerank, truncated to 2000 chars per the jina-reranker
n_ctx_train=1024 batch-rejection gotcha). Fails open to base order
on any exception.

## Baseline numbers (k=5, pool=50, 35 queries)

  | Retriever  | MRR   | Recall@5 | nDCG@5 |
  |------------|-------|----------|--------|
  | dense      | 0.027 | 0.086    | 0.041  |
  | bm25       | 0.544 | 0.586    | 0.524  |
  | hybrid-rrf | 0.114 | 0.114    | 0.108  |

Headline: BM25 dominates because farmers search for products by
brand name, and brand names are exact-match tokens that lexical
search nails. Dense is poor — semantic embeddings spread across
similar products and don't preferentially weight brand-name tokens.
Textbook RRF hurts when one retriever is much weaker than the
other: dense's irrelevant top-50 pollute the fused pool with
ties at 1/(60+rank). Phase 6 reranker is the planned fix —
the reranker scores each (query, chunk) pair independently
and can recover the right answer regardless of base order.

Per-query report at eval/results/baseline.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-24 10:19:05 -04:00
parent 97a2a05b24
commit 335c33465b
6 changed files with 636 additions and 101 deletions
+261
View File
@@ -0,0 +1,261 @@
"""Retriever protocol + concrete implementations for the labels corpus.
A single matrix dimension per knob (dense / reranked / bm25 / hybrid)
so the eval harness can compare them apples-to-apples.
Each retriever returns a ranked list of ``(source, source_key)`` tuples
deduplicated to the label level (chunks within the same label collapse
to one entry; the highest-ranked chunk's position wins). The page-level
view matches how MCP consumers think — "give me the right label" not
"give me the right chunk".
"""
from __future__ import annotations
import logging
import os
import sqlite3
from pathlib import Path
from typing import Iterable, Protocol
log = logging.getLogger(__name__)
REPO_ROOT = Path(__file__).resolve().parent.parent
CHROMA_DIR = Path(os.environ.get("PPLS_CHROMA_DIR") or REPO_ROOT / "chroma")
BM25_DB = Path(os.environ.get("BM25_DB",
str(REPO_ROOT / "bm25" / "ppls_docs.db")))
COLLECTION = f"{os.environ.get('PRODUCT_NAME', 'ppls')}_docs"
class Retriever(Protocol):
name: str
def retrieve(self, query: str, k: int = 10) -> list[tuple[str, str]]:
"""Return up to k (source, source_key) tuples in rank order."""
...
def _collapse_chunks_to_labels(
ranked_chunks: Iterable[tuple[str, str, int]], k: int
) -> list[tuple[str, str]]:
"""Stream of (source, source_key, ordinal) → top-k unique (source, source_key)
in first-seen order."""
seen: set[tuple[str, str]] = set()
out: list[tuple[str, str]] = []
for source, source_key, _ord in ranked_chunks:
key = (source, source_key)
if key in seen:
continue
seen.add(key)
out.append(key)
if len(out) >= k:
break
return out
def _parse_chunk_id(chunk_id: str) -> tuple[str, str, int]:
"""Chunk IDs look like 'source::source_key::ordinal'. Robust to
source_keys that contain '::' (none do today, but be defensive)."""
parts = chunk_id.rsplit("::", 2)
if len(parts) != 3:
return ("", chunk_id, 0)
source, source_key, ord_str = parts
try:
ord_int = int(ord_str)
except ValueError:
ord_int = 0
return (source, source_key, ord_int)
# ---------------------------------------------------------------------------
# Dense (Chroma) retriever
# ---------------------------------------------------------------------------
class DenseRetriever:
name = "dense"
def __init__(self, collection=None, over_fetch_factor: int = 4):
self.over_fetch_factor = over_fetch_factor
self._col = collection
def _collection(self):
if self._col is not None:
return self._col
import chromadb
from chromadb.config import Settings
from rag.embeddings import embedding_function
client = chromadb.PersistentClient(
path=str(CHROMA_DIR),
settings=Settings(anonymized_telemetry=False),
)
self._col = client.get_collection(
COLLECTION, embedding_function=embedding_function()
)
return self._col
def retrieve(self, query: str, k: int = 10) -> list[tuple[str, str]]:
col = self._collection()
n_fetch = max(k * self.over_fetch_factor, k)
res = col.query(query_texts=[query], n_results=n_fetch)
ids = res.get("ids", [[]])[0]
ranked: list[tuple[str, str, int]] = []
for cid in ids:
ranked.append(_parse_chunk_id(cid))
return _collapse_chunks_to_labels(ranked, k)
# ---------------------------------------------------------------------------
# BM25 (SQLite FTS5) retriever
# ---------------------------------------------------------------------------
class BM25Retriever:
"""Wraps ``rag.bm25.BM25Index`` so eval/server can call .retrieve()
on it the same way as the dense retriever. The index itself handles
FTS5 query sanitization + OR-of-tokens semantics."""
name = "bm25"
def __init__(self, db_path: Path = BM25_DB, over_fetch_factor: int = 4):
from rag.bm25 import BM25Index
self._idx = BM25Index(db_path)
self.over_fetch_factor = over_fetch_factor
def retrieve(self, query: str, k: int = 10) -> list[tuple[str, str]]:
n_fetch = max(k * self.over_fetch_factor, k)
hits = self._idx.query(query, n=n_fetch)
ranked = [_parse_chunk_id(cid) for cid, _score in hits]
return _collapse_chunks_to_labels(ranked, k)
# ---------------------------------------------------------------------------
# Hybrid retriever (BM25 + dense, RRF fusion)
# ---------------------------------------------------------------------------
class HybridRetriever:
"""Reciprocal Rank Fusion of dense + BM25 results. The fused score
for a page p is sum over retrievers r of 1 / (k_rrf + rank_r(p)).
Pages absent from a retriever contribute 0 from it."""
name = "hybrid-rrf"
def __init__(
self,
dense: DenseRetriever | None = None,
bm25: BM25Retriever | None = None,
k_rrf: int = 60,
pool: int = 50,
):
self.dense = dense or DenseRetriever()
self.bm25 = bm25 or BM25Retriever()
self.k_rrf = k_rrf
self.pool = pool
def retrieve(self, query: str, k: int = 10) -> list[tuple[str, str]]:
dense_pages = self.dense.retrieve(query, k=self.pool)
bm25_pages = self.bm25.retrieve(query, k=self.pool)
scores: dict[tuple[str, str], float] = {}
for rank, page in enumerate(dense_pages, start=1):
scores[page] = scores.get(page, 0.0) + 1.0 / (self.k_rrf + rank)
for rank, page in enumerate(bm25_pages, start=1):
scores[page] = scores.get(page, 0.0) + 1.0 / (self.k_rrf + rank)
fused = sorted(scores.items(), key=lambda kv: -kv[1])
return [page for page, _ in fused[:k]]
# ---------------------------------------------------------------------------
# Reranker (jina-reranker via llama.cpp /v1/rerank)
# ---------------------------------------------------------------------------
class RerankedRetriever:
"""Take a base retriever's pool, fetch full chunk text for each page's
top chunk, send (query, chunk_text) pairs to a llama.cpp /v1/rerank
endpoint, then rerank pages by the returned scores.
For eval we operate page-level. We pick the first chunk per page from
the base retriever's chunk-level output. To get the chunk text we
re-query Chroma by chunk id."""
name = "dense+rerank"
def __init__(
self,
base: Retriever | None = None,
rerank_url: str | None = None,
pool: int = 50,
timeout: float = 30.0,
):
self.base = base or DenseRetriever()
self.rerank_url = (rerank_url or os.environ.get("RERANK_URL", "")).rstrip("/")
self.pool = pool
self.timeout = timeout
self._col = None
@property
def name_with_base(self) -> str:
return f"{self.base.name}+rerank"
def _collection(self):
if self._col is not None:
return self._col
import chromadb
from chromadb.config import Settings
client = chromadb.PersistentClient(
path=str(CHROMA_DIR),
settings=Settings(anonymized_telemetry=False),
)
# We don't need the embedder for fetch-by-id; pass embedding_function=None
self._col = client.get_collection(COLLECTION)
return self._col
def retrieve(self, query: str, k: int = 10) -> list[tuple[str, str]]:
if not self.rerank_url:
# Fail open to base retriever — useful in eval to compare base vs
# base+rerank when the reranker is offline.
log.warning("RERANK_URL unset; falling back to base retriever")
return self.base.retrieve(query, k=k)
pages = self.base.retrieve(query, k=self.pool)
if not pages:
return []
# Fetch one representative chunk per page (the first chunk, ordinal=0
# if it exists, else any). For eval simplicity we approximate by
# fetching by metadata where (source, source_key) and taking the
# first hit.
col = self._collection()
docs: list[str] = []
kept_pages: list[tuple[str, str]] = []
for source, source_key in pages:
where = {"$and": [{"source": source}, {"source_key": source_key}]}
got = col.get(where=where, limit=1, include=["documents"])
d = (got.get("documents") or [None])[0]
if not d:
continue
# Truncate to keep under the reranker's per-pair context limit
docs.append(d[:2000])
kept_pages.append((source, source_key))
if not docs:
return []
import httpx
try:
r = httpx.post(
f"{self.rerank_url}/v1/rerank",
json={"query": query, "documents": docs},
timeout=self.timeout,
)
r.raise_for_status()
data = r.json()
except Exception as exc: # noqa: BLE001
log.warning("rerank failed (%s) — falling back to base order", exc)
return kept_pages[:k]
# llama.cpp returns {"results": [{"index": i, "relevance_score": s}, ...]}
results = data.get("results") or []
scored: list[tuple[float, tuple[str, str]]] = []
for r_item in results:
idx = r_item.get("index")
score = r_item.get("relevance_score") or r_item.get("score") or 0.0
if isinstance(idx, int) and 0 <= idx < len(kept_pages):
scored.append((score, kept_pages[idx]))
scored.sort(key=lambda x: -x[0])
return [p for _, p in scored[:k]]