Angshul's picture
Upload 18 files
f7b6133 verified
Raw
History Blame Contribute Delete
14.6 kB
from __future__ import annotations
"""Top-10 RAG ranking layer for the sparse geometric index.
This module is the cleaned, path-independent version of the exact SciFact and
TREC-COVID experiment scripts preserved under ``experiments/beir/*history.py``.
It keeps the geometric index fixed and changes only the shortlist size P and
the final top-10 set construction.
Important implementation choices
--------------------------------
* Early rescue: binary whole-chunk IDF^1 support.
* Final lexical signal: binary whole-chunk IDF^2 support, not TF^2.
* Final components retain the validated per-query z-normalization.
* Branch quality H_j is the mean of the top three branch-specific evidences.
* Diversity is available only to the ten highest-quality branches.
* Rank 1 is pure relevance; ranks 2..10 receive a soft diversity correction.
* Repeated branches are allowed. There is no one-document-per-branch rule.
"""
from dataclasses import dataclass
import time
import numpy as np
from .metrics import evaluate_run
def _zscore(x):
x = np.asarray(x, np.float32)
if not len(x):
return x
sd = float(x.std())
return np.zeros_like(x) if sd < 1e-8 else (x - float(x.mean())) / sd
def _minmax_hi(x):
"""Map scores monotonically to [0, 1], high remains good."""
x = np.asarray(x, np.float32)
if not len(x):
return x
lo, hi = float(x.min()), float(x.max())
den = hi - lo
return np.ones_like(x) if den < 1e-8 else (x - lo) / den
def _topk_large(score, k):
score = np.asarray(score)
n = len(score)
k = min(int(k), n)
if k <= 0:
return np.empty(0, np.int64)
if n <= k:
return np.argsort(score)[::-1]
ii = np.argpartition(score, -k)[-k:]
return ii[np.argsort(score[ii])[::-1]]
@dataclass(frozen=True)
class RAGTop10Config:
pool_size: int = 100
gamma: float = 0.25
lambda_membership: float = 0.125
pre_length_b: float = 0.2
final_length_b: float = 0.1
coordination_alpha: float = 0.25
lambda_lex: float = 4.0
lambda_sem: float = 0.3
lambda_rare: float = 1.0
semantic_k: int = 16
rare_topk: int = 3
hq_top_branches: int = 10
branch_quality_top_docs: int = 3
lambda_diversity: float = 0.1
class RAGTop10Ranker:
"""RAG-oriented shortlist and top-10 set selector.
``GeometricIndex`` performs corpus indexing and stores the compact geometry.
This class consumes that frozen representation. It does not train a model,
rebuild centers, or alter residual codes.
"""
def __init__(self, index, config: RAGTop10Config | None = None):
self.idx = index
self.cfg = config or RAGTop10Config()
self.M = int(index.vocab_size)
def _center_sparse(self, branch):
t = self.idx.center_terms[branch]
v = self.idx.center_values[branch]
ok = t >= 0
t = t[ok].astype(np.int32)
v = v[ok].astype(np.float32)
n = float(np.linalg.norm(v))
if n > 0:
v = v / n
order = np.argsort(t)
return t[order], v[order]
@staticmethod
def _spdot(a_t, a_v, b_t, b_v):
i = j = 0
s = 0.0
while i < len(a_t) and j < len(b_t):
if a_t[i] == b_t[j]:
s += float(a_v[i]) * float(b_v[j]); i += 1; j += 1
elif a_t[i] < b_t[j]:
i += 1
else:
j += 1
return s
def prepare(self, text: str):
"""Retrieve geometric candidates and create the P-sized chunk shortlist."""
idx, cfg, M = self.idx, self.cfg, self.M
q = idx._query_vector(text)
if q.nnz == 0:
return None
q_dense = np.zeros(M, np.float32)
q_dense[q.indices] = q.data
route_terms, route_values, _ = idx._expanded_route(q)
if not len(route_terms):
return None
route_dense = np.zeros(M, np.float32)
route_dense[route_terms] = route_values
pieces = []
for j in route_terms:
a, b = idx.branch_offsets[j], idx.branch_offsets[j + 1]
if b > a:
pieces.append(idx.branch_order[a:b])
if not pieces:
return None
flat_pos = np.concatenate(pieces).astype(np.int64, copy=False)
docs = (flat_pos // idx.config.F).astype(np.int64)
slots = (flat_pos % idx.config.F).astype(np.int64)
branches = idx.branches[docs, slots]
terms = idx.res_terms[docs, slots]
valid = terms >= 0
safe_terms = np.where(valid, terms, 0)
qv = q_dense[safe_terms]
local = np.sum(
idx.res_reliability[docs, slots].astype(np.float32)
* (qv - idx.res_center_values[docs, slots])
* idx.res_signs[docs, slots].astype(np.float32)
* valid,
axis=1,
)
significance = np.sum((qv * qv) * valid, axis=1)
consensus = idx.memberships[docs, slots] * route_dense[branches]
branch_ev = (
consensus * local * np.power(np.maximum(significance, 0), cfg.gamma)
).astype(np.float32)
unique_docs, inverse = np.unique(docs, return_inverse=True)
tail = np.bincount(inverse, weights=branch_ev, minlength=len(unique_docs)).astype(np.float32)
tail += cfg.lambda_membership * np.bincount(
inverse, weights=consensus, minlength=len(unique_docs)
).astype(np.float32)
# Stage 1: cheap whole-chunk binary IDF^1 rescue before expensive final scoring.
qlex = np.zeros(M, np.float32)
qlex[q.indices] = idx.idf[q.indices]
lex1 = np.zeros(len(unique_docs), np.float32)
for i, d in enumerate(unique_docs):
a, b = idx.support_indptr[d], idx.support_indptr[d + 1]
support = idx.support_indices[a:b]
raw = float(qlex[support].sum())
den = (1 - cfg.pre_length_b) + cfg.pre_length_b * (
float(idx.doc_lengths[d]) / idx.avg_doc_length
)
lex1[i] = raw / (den if den > 0 else 1.0)
pre = _zscore(tail) + _zscore(lex1)
selected = _topk_large(pre, cfg.pool_size)
pool_docs = unique_docs[selected]
pool_tail = tail[selected]
# Preserve branch-specific evidence for robust branch-quality estimation.
pool_position = np.full(len(unique_docs), -1, np.int32)
pool_position[selected] = np.arange(len(selected), dtype=np.int32)
mapped = pool_position[inverse]
keep = mapped >= 0
mem_pool = mapped[keep].astype(np.int32)
mem_branch = branches[keep].astype(np.int32)
mem_ev = branch_ev[keep].astype(np.float32)
# Stage 2: final chunk evidence. No document TF is used here.
semvec = np.zeros(M, np.float32)
for t, amp in zip(q.indices, q.data):
a, b = idx.A.indptr[t], idx.A.indptr[t + 1]
nb = idx.A.indices[a:b][: cfg.semantic_k]
sv = idx.A.data[a:b][: cfg.semantic_k]
if len(nb):
semvec[nb] += float(amp) * sv * idx.idf[nb]
qset = set(map(int, q.indices))
rare = set(map(int, q.indices[np.argsort(idx.idf[q.indices])[::-1]][: cfg.rare_topk]))
nq = max(1, len(q.indices))
lex2 = np.zeros(len(pool_docs), np.float32)
sem = np.zeros(len(pool_docs), np.float32)
matched_count = np.zeros(len(pool_docs), np.float32)
rare_count = np.zeros(len(pool_docs), np.float32)
for i, d in enumerate(pool_docs):
a, b = idx.support_indptr[d], idx.support_indptr[d + 1]
support = idx.support_indices[a:b]
sem[i] = float(semvec[support].sum())
match = [int(t) for t in support if int(t) in qset]
raw = sum(float(idx.idf[t]) ** 2 for t in match)
den = (1 - cfg.final_length_b) + cfg.final_length_b * (
float(idx.doc_lengths[d]) / idx.avg_doc_length
)
lex2[i] = raw / (den if den > 0 else 1.0)
matched_count[i] = len(match)
rare_count[i] = sum(t in rare for t in match)
coverage = matched_count / nq
lex_adjusted = lex2 * np.power(np.maximum(coverage, 1e-6), cfg.coordination_alpha)
rare_coverage = rare_count / max(1, min(cfg.rare_topk, nq))
relevance = (
_zscore(pool_tail)
+ cfg.lambda_lex * _zscore(lex_adjusted)
+ cfg.lambda_sem * _zscore(sem)
+ cfg.lambda_rare * _zscore(rare_coverage)
)
# Robust high-quality branch score H_j: mean top-r branch-specific evidence.
branch_pairs = {}
for pi, b, e in zip(mem_pool, mem_branch, mem_ev):
key = (int(b), int(pi))
if key not in branch_pairs or e > branch_pairs[key]:
branch_pairs[key] = float(e)
by_branch = {}
for (b, pi), e in branch_pairs.items():
by_branch.setdefault(b, []).append((e, pi))
H, docs_by_branch = {}, {}
for b, vals in by_branch.items():
vals.sort(key=lambda x: x[0], reverse=True)
r = min(cfg.branch_quality_top_docs, len(vals))
H[b] = float(np.mean([e for e, _ in vals[:r]]))
docs_by_branch[b] = np.asarray([pi for _, pi in vals], dtype=np.int32)
unique_branches = np.asarray(sorted(H.keys()), dtype=np.int32)
h = np.asarray([H[int(b)] for b in unique_branches], dtype=np.float32)
centers = [self._center_sparse(int(b)) for b in unique_branches]
cosine = np.eye(len(unique_branches), dtype=np.float32)
for i in range(len(unique_branches)):
for j in range(i + 1, len(unique_branches)):
cosine[i, j] = cosine[j, i] = self._spdot(*centers[i], *centers[j])
return {
"docs": pool_docs,
"relevance": relevance,
"route_docs": unique_docs,
"branches": unique_branches,
"branch_quality": h,
"cosine": cosine,
"docs_by_branch": docs_by_branch,
}
@staticmethod
def _deviation(cosine, selected_branch_indices):
"""Squared distance of each unit branch center from selected-center centroid."""
if not selected_branch_indices:
return np.zeros(len(cosine), np.float32)
si = np.asarray(selected_branch_indices, np.int32)
centroid_norm_sq = float(np.mean(cosine[np.ix_(si, si)]))
return 1.0 + centroid_norm_sq - 2.0 * np.mean(cosine[:, si], axis=1)
def rank(self, packet, k: int = 10):
"""Construct top-k; branch diversity affects only the first ten ranks."""
if packet is None or not len(packet["docs"]):
return []
cfg = self.cfg
base = packet["relevance"]
n = len(base)
order = np.argsort(base)[::-1]
first = int(order[0])
chosen = [first]
used = {first}
# Only top query-specific branches are eligible to receive a diversity bonus.
h = packet["branch_quality"]
eligible = np.argsort(h)[::-1][: min(cfg.hq_top_branches, len(h))]
doc_hq = [[] for _ in range(n)]
branch_to_index = {int(b): i for i, b in enumerate(packet["branches"])}
for bi in eligible:
b = int(packet["branches"][bi])
for pi in packet["docs_by_branch"].get(b, []):
doc_hq[int(pi)].append(int(bi))
selected_branches = []
if doc_hq[first]:
selected_branches = [max(doc_hq[first], key=lambda bi: float(h[bi]))]
# Ranks 2..10: relevance plus soft central-deviation bonus.
for _ in range(1, min(10, k, n)):
rem = np.asarray([i for i in range(n) if i not in used], dtype=np.int32)
if not len(rem):
break
dev = self._deviation(packet["cosine"], selected_branches)
if len(eligible):
dev_values = _minmax_hi(dev[eligible])
dev_map = {int(bi): float(v) for bi, v in zip(eligible, dev_values)}
else:
dev_map = {}
rel = _minmax_hi(base[rem])
bonus = np.zeros(len(rem), np.float32)
for ii, pi in enumerate(rem):
if doc_hq[int(pi)]:
bonus[ii] = max(dev_map.get(bi, 0.0) for bi in doc_hq[int(pi)])
value = rel + cfg.lambda_diversity * bonus
pi = int(rem[int(np.argmax(value))])
chosen.append(pi)
used.add(pi)
if doc_hq[pi]:
bi = max(doc_hq[pi], key=lambda x: (dev_map.get(x, 0.0), float(h[x])))
selected_branches.append(int(bi))
# Beyond rank 10, ordinary relevance. This keeps the top-10 RAG mechanism isolated.
for pi in order:
pi = int(pi)
if len(chosen) >= min(k, n):
break
if pi not in used:
chosen.append(pi)
used.add(pi)
return packet["docs"][np.asarray(chosen[:k], dtype=np.int64)].tolist()
def search(self, text: str, k: int = 10, timing: bool = False):
t0 = time.perf_counter()
packet = self.prepare(text)
t1 = time.perf_counter()
local_ids = self.rank(packet, k=k)
t2 = time.perf_counter()
doc_ids = [str(self.idx.doc_ids[int(d)]) for d in local_ids]
if not timing:
return doc_ids
return doc_ids, {
"prepare_ms": (t1 - t0) * 1000.0,
"rank_ms": (t2 - t1) * 1000.0,
"total_ms": (t2 - t0) * 1000.0,
"route_size": 0 if packet is None else len(packet["route_docs"]),
"pool_size": 0 if packet is None else len(packet["docs"]),
}
def evaluate(self, dataset, k: int = 10):
run = {}
times = []
for qid in dataset.qrels:
docs, timing = self.search(dataset.queries[qid], k=k, timing=True)
run[str(qid)] = docs
times.append(timing)
metrics = evaluate_run(run, dataset.qrels, ks=(10,), ndcg_k=10, mrr_k=10, exp_gain=False)
metrics.update({
"median_total_ms": float(np.median([x["total_ms"] for x in times])),
"p95_total_ms": float(np.percentile([x["total_ms"] for x in times], 95)),
"median_rank_ms": float(np.median([x["rank_ms"] for x in times])),
"median_route_size": float(np.median([x["route_size"] for x in times])),
"median_pool_size": float(np.median([x["pool_size"] for x in times])),
})
return metrics, run