ZacharyZhang-NY
a4953335e7
The merged database names a source by content hash (`11089_4f2a...`) while the frozen sets name the same file by its corpus id, so every arm scored 0.000 until load_atoms joined atoms through sources.path. It now returns corpus ids and refuses an atom it cannot trace to a corpus file. Also fixes found by running the harness: - merge rebuilt atom_fts from the deduped atoms table; two attempts on one file left 288,266 fts rows for 165,973 atoms. - merge accepts the two honest reasons a corpus file has no product source (500 byte-identical LongMemEval sessions collapse under content hashing, and 12 photos that ONNX Runtime cannot decode) and reports them. - --arms used action=extend against a non-empty default, so naming two arms ran seven. - the dense cache is keyed by the atom set, so a scale-curve subset cannot be scored with full-corpus vectors. - write_markdown crashed on a run that covered only some modalities.
322 lines
13 KiB
Python
322 lines
13 KiB
Python
"""Retrieval arms for the ablation. Every arm reads the same merged baseline database.
|
|
|
|
The arms are the product's own signals, switched on one at a time: sparse BM25, dense
|
|
cosine, the evidence graph, the product's weighted fusion, and fusion plus a cross-encoder
|
|
reranker. Rankings are lists of source ids: an atom votes for its source, and a source is
|
|
ranked by its best atom.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import hashlib
|
|
import json
|
|
import re
|
|
import sqlite3
|
|
import time
|
|
from collections.abc import Sequence
|
|
from dataclasses import dataclass
|
|
from pathlib import Path
|
|
from typing import Any
|
|
|
|
import numpy as np
|
|
|
|
from bench.ingest_baseline import source_map
|
|
from bench.sets_common import ROOT
|
|
|
|
RUNS = ROOT / "runs" / "baseline"
|
|
INDEX_DIR = RUNS / "index"
|
|
MODELS = ROOT / "models"
|
|
DENSE_MODEL = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
|
|
RERANK_MODEL = "BAAI/bge-reranker-base"
|
|
|
|
|
|
def local_model_dir(models: Path, repo_id: str) -> Path:
|
|
"""Where a fetched model lives: one flat directory per repo, named by its repo name only."""
|
|
return models / repo_id.split("/")[-1]
|
|
DENSE_DIM = 384
|
|
FUSION_WEIGHTS = {"sparse": 0.55, "dense": 0.35, "graph": 0.10}
|
|
RERANK_CANDIDATES = 50
|
|
RERANK_TEXT_CHARS = 400
|
|
ARMS = ("bm25_only", "dense_only", "graph_only", "no_rerank", "fusion_rerank")
|
|
|
|
|
|
@dataclass(frozen=True)
|
|
class Atom:
|
|
atom_id: str
|
|
source_id: str
|
|
text: str
|
|
|
|
|
|
def load_atoms(db_path: Path) -> list[Atom]:
|
|
"""Every atom in the merged database, keyed by the benchmark's source id.
|
|
|
|
The product names a source by content hash (`11089_4f2a...`), while the frozen
|
|
sets name the same file by its corpus id. `sources.path` is the only column
|
|
that identifies the original file, so the two id spaces are joined through it.
|
|
"""
|
|
if not db_path.is_file():
|
|
raise SystemExit(f"{db_path}: no merged baseline database, run python -m bench.ingest_baseline first")
|
|
connection = sqlite3.connect(f"file:{db_path}?mode=ro", uri=True)
|
|
try:
|
|
rows = connection.execute(
|
|
"SELECT a.atom_id, s.path, a.text FROM atoms AS a "
|
|
"JOIN sources AS s ON s.source_id = a.source_id ORDER BY a.atom_id"
|
|
).fetchall()
|
|
except sqlite3.OperationalError as error:
|
|
raise SystemExit(f"{db_path}: cannot read the atoms table ({error})") from error
|
|
finally:
|
|
connection.close()
|
|
if not rows:
|
|
raise SystemExit(f"{db_path}: the atoms table is empty")
|
|
mapping = source_map()
|
|
unknown = sorted({str(path) for _, path, _ in rows if str(path) not in mapping})
|
|
if unknown:
|
|
raise SystemExit(f"{db_path}: {len(unknown)} atom paths are not corpus files, first {unknown[:3]}")
|
|
return [
|
|
Atom(atom_id=str(atom_id), source_id=str(mapping[str(path)]), text=str(text or ""))
|
|
for atom_id, path, text in rows
|
|
]
|
|
|
|
|
|
def tokenize(text: str) -> list[str]:
|
|
"""The product's tokenizer: word characters plus CJK, lowercased."""
|
|
return re.findall(r"[\w\u4e00-\u9fff]+", text.lower())
|
|
|
|
|
|
class SparseIndex:
|
|
"""BM25 over atom text, built once; the product rebuilds it on every query."""
|
|
|
|
def __init__(self, atoms: Sequence[Atom]) -> None:
|
|
from rank_bm25 import BM25Okapi
|
|
|
|
corpus = [tokenize(f"{atom.text} {atom.source_id}") for atom in atoms]
|
|
self._model = BM25Okapi(corpus)
|
|
self._count = len(atoms)
|
|
|
|
def scores(self, query: str) -> np.ndarray:
|
|
tokens = tokenize(query)
|
|
if not tokens:
|
|
return np.zeros(self._count, dtype=np.float64)
|
|
return np.asarray(self._model.get_scores(tokens), dtype=np.float64)
|
|
|
|
|
|
def _cache_key(model_dir: Path, atoms: Sequence[Atom]) -> str:
|
|
"""A cache name that follows the atom set, so a scale-curve subset can never
|
|
be scored with the full corpus's vectors (or the other way round)."""
|
|
joined = "\n".join(atom.atom_id for atom in atoms)
|
|
return f"{model_dir.name}-{hashlib.sha256(joined.encode('utf-8')).hexdigest()[:16]}"
|
|
|
|
|
|
class DenseIndex:
|
|
"""Brute-force cosine over cached sentence embeddings, one vector per atom."""
|
|
|
|
def __init__(self, atoms: Sequence[Atom], model_dir: Path, batch_size: int = 256, cache_name: str = "") -> None:
|
|
self._cache_name = cache_name or _cache_key(model_dir, atoms)
|
|
self._vectors = self._load_or_build(atoms, model_dir, batch_size)
|
|
|
|
def _vectors_path(self, model_dir: Path) -> Path:
|
|
return INDEX_DIR / f"{self._cache_name}.npy"
|
|
|
|
def _load_or_build(self, atoms: Sequence[Atom], model_dir: Path, batch_size: int) -> np.ndarray:
|
|
path = self._vectors_path(model_dir)
|
|
order_path = INDEX_DIR / f"{self._cache_name}.atoms.json"
|
|
atom_ids = [atom.atom_id for atom in atoms]
|
|
if path.is_file() and order_path.is_file():
|
|
try:
|
|
vectors = np.load(path)
|
|
cached_ids = json.loads(order_path.read_text(encoding="utf-8"))
|
|
except (OSError, ValueError) as error:
|
|
raise SystemExit(f"{order_path}: cached dense index is unreadable ({error})") from error
|
|
if cached_ids == atom_ids and vectors.shape == (len(atoms), DENSE_DIM):
|
|
return vectors
|
|
if cached_ids != atom_ids:
|
|
raise SystemExit(
|
|
f"{path}: cached vectors cover {len(cached_ids)} atoms but the database has {len(atoms)}; "
|
|
"delete bench/runs/baseline/index and rebuild"
|
|
)
|
|
INDEX_DIR.mkdir(parents=True, exist_ok=True)
|
|
model = _sentence_model(model_dir)
|
|
texts = [atom.text[:RERANK_TEXT_CHARS] if atom.text else atom.source_id for atom in atoms]
|
|
vectors = model.encode(
|
|
texts,
|
|
batch_size=batch_size,
|
|
convert_to_numpy=True,
|
|
normalize_embeddings=True,
|
|
show_progress_bar=len(texts) > 2000,
|
|
).astype(np.float32)
|
|
if vectors.shape[1] != DENSE_DIM:
|
|
raise SystemExit(f"{model_dir}: expected {DENSE_DIM} dimensions, got {vectors.shape[1]}")
|
|
np.save(path, vectors)
|
|
order_path.write_text(json.dumps(atom_ids), encoding="utf-8")
|
|
return vectors
|
|
|
|
def scores(self, query: str, model_dir: Path) -> np.ndarray:
|
|
model = _sentence_model(model_dir)
|
|
vector = model.encode([query], normalize_embeddings=True, convert_to_numpy=True).astype(np.float32)
|
|
return np.asarray(self._vectors @ vector[0], dtype=np.float64)
|
|
|
|
|
|
def _sentence_model(model_dir: Path) -> Any:
|
|
"""sentence-transformers ships no stubs, so the model is typed Any rather than a fake object."""
|
|
from sentence_transformers import SentenceTransformer
|
|
|
|
if not model_dir.is_dir():
|
|
raise SystemExit(f"{model_dir}: model not downloaded, run python -m bench.fetch_models first")
|
|
return SentenceTransformer(str(model_dir), local_files_only=True)
|
|
|
|
|
|
def top_sources(scores: np.ndarray, atoms: Sequence[Atom], top_k: int) -> list[str]:
|
|
"""Rank sources by their best-scoring atom; ties break on source id."""
|
|
if len(scores) != len(atoms):
|
|
raise SystemExit(f"scores has {len(scores)} entries for {len(atoms)} atoms")
|
|
best: dict[str, float] = {}
|
|
for atom, score in zip(atoms, scores, strict=True):
|
|
current = best.get(atom.source_id)
|
|
if current is None or score > current:
|
|
best[atom.source_id] = float(score)
|
|
ranked = sorted(best.items(), key=lambda item: (-item[1], item[0]))
|
|
return [source_id for source_id, _ in ranked[:top_k]]
|
|
|
|
|
|
def fuse(sparse: np.ndarray, dense: np.ndarray, graph: np.ndarray) -> np.ndarray:
|
|
"""The product's weighted mix, with each signal scaled to [0, 1]."""
|
|
return (
|
|
FUSION_WEIGHTS["sparse"] * _unit(sparse)
|
|
+ FUSION_WEIGHTS["dense"] * _unit(dense)
|
|
+ FUSION_WEIGHTS["graph"] * _unit(graph)
|
|
)
|
|
|
|
|
|
def _unit(values: np.ndarray) -> np.ndarray:
|
|
if values.size == 0:
|
|
return values
|
|
low = float(values.min())
|
|
high = float(values.max())
|
|
if high - low <= 1e-12:
|
|
return np.zeros_like(values)
|
|
return (values - low) / (high - low)
|
|
|
|
|
|
class Reranker:
|
|
"""Cross-encoder over the fusion candidate list."""
|
|
|
|
def __init__(self, model_dir: Path) -> None:
|
|
if not model_dir.is_dir():
|
|
raise SystemExit(f"{model_dir}: model not downloaded, run python -m bench.fetch_models first")
|
|
from sentence_transformers import CrossEncoder
|
|
|
|
self._model = CrossEncoder(str(model_dir), local_files_only=True)
|
|
|
|
def scores(self, query: str, texts: Sequence[str]) -> np.ndarray:
|
|
if not texts:
|
|
return np.zeros(0, dtype=np.float64)
|
|
pairs = [(query, text[:RERANK_TEXT_CHARS] or " ") for text in texts]
|
|
return np.asarray(self._model.predict(pairs, batch_size=16), dtype=np.float64)
|
|
|
|
|
|
def arm_needs(arm: str, models: Path) -> bool:
|
|
"""Whether an arm's model files are present, so a run can skip it instead of crashing."""
|
|
if arm not in ARMS:
|
|
raise SystemExit(f"{arm}: unknown arm, expected one of {ARMS}")
|
|
required: list[Path] = []
|
|
if arm in ("dense_only", "no_rerank", "fusion_rerank"):
|
|
required.append(local_model_dir(models, DENSE_MODEL))
|
|
if arm == "fusion_rerank":
|
|
required.append(local_model_dir(models, RERANK_MODEL))
|
|
return all(directory.is_dir() for directory in required)
|
|
|
|
|
|
class ArmRunner:
|
|
"""One arm, ranked queries in batches, with wall-clock timings recorded."""
|
|
|
|
def __init__(self, arm: str, atoms: Sequence[Atom], models: Path) -> None:
|
|
if arm not in ARMS:
|
|
raise SystemExit(f"{arm}: unknown arm, expected one of {ARMS}")
|
|
self.arm = arm
|
|
self.atoms = atoms
|
|
self.models = models
|
|
self.timings: list[float] = []
|
|
self._sparse: SparseIndex | None = None
|
|
self._dense: DenseIndex | None = None
|
|
self._reranker: Reranker | None = None
|
|
if arm in ("bm25_only", "no_rerank", "fusion_rerank"):
|
|
self._sparse = SparseIndex(atoms)
|
|
if arm in ("dense_only", "no_rerank", "fusion_rerank"):
|
|
self._dense = DenseIndex(atoms, local_model_dir(models, DENSE_MODEL))
|
|
if arm == "fusion_rerank":
|
|
self._reranker = Reranker(local_model_dir(models, RERANK_MODEL))
|
|
|
|
def rank(self, query: str, top_k: int = 10) -> list[str]:
|
|
started = time.perf_counter()
|
|
ranked = self._rank(query, top_k)
|
|
self.timings.append(time.perf_counter() - started)
|
|
return ranked
|
|
|
|
def _rank(self, query: str, top_k: int) -> list[str]:
|
|
if self.arm == "graph_only":
|
|
# The product never populates atom.relations, so the graph signal is a constant.
|
|
graph = np.zeros(len(self.atoms), dtype=np.float64)
|
|
return top_sources(graph, self.atoms, top_k)
|
|
sparse = self._sparse.scores(query) if self._sparse is not None else np.zeros(len(self.atoms))
|
|
if self.arm == "bm25_only":
|
|
return top_sources(sparse, self.atoms, top_k)
|
|
dense_dir = local_model_dir(self.models, DENSE_MODEL)
|
|
dense = (
|
|
self._dense.scores(query, dense_dir) if self._dense is not None else np.zeros(len(self.atoms))
|
|
)
|
|
if self.arm == "dense_only":
|
|
return top_sources(dense, self.atoms, top_k)
|
|
graph = np.zeros(len(self.atoms), dtype=np.float64)
|
|
fused = fuse(sparse, dense, graph)
|
|
if self.arm == "no_rerank":
|
|
return top_sources(fused, self.atoms, top_k)
|
|
candidates = top_sources(fused, self.atoms, RERANK_CANDIDATES)
|
|
by_source = _best_atom_text(self.atoms, candidates)
|
|
if self._reranker is None:
|
|
raise SystemExit(f"{self.arm}: the reranker model is missing, run python -m bench.fetch_models")
|
|
rerank_scores = self._reranker.scores(query, [text for _, text in by_source])
|
|
order = sorted(zip(candidates, rerank_scores, strict=True), key=lambda item: (-item[1], item[0]))
|
|
return [source_id for source_id, _ in order[:top_k]]
|
|
|
|
def percentile(self, fraction: float) -> float:
|
|
if not self.timings:
|
|
raise SystemExit(f"{self.arm}: no timings recorded")
|
|
if not 0.0 <= fraction <= 1.0:
|
|
raise SystemExit(f"percentile must be within [0, 1], got {fraction}")
|
|
return float(np.percentile(np.asarray(self.timings), fraction * 100))
|
|
|
|
|
|
def _best_atom_text(atoms: Sequence[Atom], source_ids: Sequence[str]) -> list[tuple[str, str]]:
|
|
longest: dict[str, str] = {}
|
|
for atom in atoms:
|
|
if atom.source_id in longest:
|
|
if len(atom.text) > len(longest[atom.source_id]):
|
|
longest[atom.source_id] = atom.text
|
|
else:
|
|
longest[atom.source_id] = atom.text
|
|
return [(source_id, longest.get(source_id, "")) for source_id in source_ids]
|
|
|
|
|
|
def live_product_rankings(db_home: Path, queries: Sequence[str], top_k: int = 10) -> tuple[list[list[str]], list[float]]:
|
|
"""Rankings from the unmodified RetrievalService, with per-query wall time."""
|
|
from ely_eye.config import Settings
|
|
from ely_eye.db import Database
|
|
from ely_eye.retrieval import RetrievalService
|
|
|
|
settings = Settings(home=db_home)
|
|
service = RetrievalService(settings=settings, database=Database(settings))
|
|
rankings: list[list[str]] = []
|
|
timings: list[float] = []
|
|
for query in queries:
|
|
started = time.perf_counter()
|
|
hits = service.search(query, top_k=top_k)
|
|
timings.append(time.perf_counter() - started)
|
|
# The product ranks atoms; collapse to sources by first (best) hit so the arms stay comparable.
|
|
sources: list[str] = []
|
|
for hit in hits:
|
|
source_id = hit.atom.source_id
|
|
if source_id not in sources:
|
|
sources.append(source_id)
|
|
rankings.append(sources[:top_k])
|
|
return rankings, timings
|