Spaces:
Running
Running
File size: 12,374 Bytes
54f5afe |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 |
"""Query documents."""
import re
import string
from collections import defaultdict
from collections.abc import Sequence
from itertools import groupby
from typing import cast
import numpy as np
from langdetect import detect
from sqlalchemy.engine import make_url
from sqlmodel import Session, and_, col, or_, select, text
from raglite._config import RAGLiteConfig
from raglite._database import Chunk, ChunkEmbedding, IndexMetadata, create_database_engine
from raglite._embed import embed_sentences
from raglite._typing import FloatMatrix
def vector_search(
query: str | FloatMatrix,
*,
num_results: int = 3,
config: RAGLiteConfig | None = None,
) -> tuple[list[str], list[float]]:
"""Search chunks using ANN vector search."""
# Read the config.
config = config or RAGLiteConfig()
db_backend = make_url(config.db_url).get_backend_name()
# Get the index metadata (including the query adapter, and in the case of SQLite, the index).
index_metadata = IndexMetadata.get("default", config=config)
# Embed the query.
query_embedding = (
embed_sentences([query], config=config)[0, :] if isinstance(query, str) else np.ravel(query)
)
# Apply the query adapter to the query embedding.
Q = index_metadata.get("query_adapter") # noqa: N806
if config.vector_search_query_adapter and Q is not None:
query_embedding = (Q @ query_embedding).astype(query_embedding.dtype)
# Search for the multi-vector chunk embeddings that are most similar to the query embedding.
if db_backend == "postgresql":
# Check that the selected metric is supported by pgvector.
metrics = {"cosine": "<=>", "dot": "<#>", "euclidean": "<->", "l1": "<+>", "l2": "<->"}
if config.vector_search_index_metric not in metrics:
error_message = f"Unsupported metric {config.vector_search_index_metric}."
raise ValueError(error_message)
# With pgvector, we can obtain the nearest neighbours and similarities with a single query.
engine = create_database_engine(config)
with Session(engine) as session:
distance_func = getattr(
ChunkEmbedding.embedding, f"{config.vector_search_index_metric}_distance"
)
distance = distance_func(query_embedding).label("distance")
results = session.exec(
select(ChunkEmbedding.chunk_id, distance).order_by(distance).limit(8 * num_results)
)
chunk_ids_, distance = zip(*results, strict=True)
chunk_ids, similarity = np.asarray(chunk_ids_), 1.0 - np.asarray(distance)
elif db_backend == "sqlite":
# Load the NNDescent index.
index = index_metadata.get("index")
ids = np.asarray(index_metadata.get("chunk_ids"))
cumsum = np.cumsum(np.asarray(index_metadata.get("chunk_sizes")))
# Find the neighbouring multi-vector indices.
from pynndescent import NNDescent
multi_vector_indices, distance = cast(NNDescent, index).query(
query_embedding[np.newaxis, :], k=8 * num_results
)
similarity = 1 - distance[0, :]
# Transform the multi-vector indices into chunk indices, and then to chunk ids.
chunk_indices = np.searchsorted(cumsum, multi_vector_indices[0, :], side="right") + 1
chunk_ids = np.asarray([ids[chunk_index - 1] for chunk_index in chunk_indices])
# Score each unique chunk id as the mean similarity of its multi-vector hits. Chunk ids with
# fewer hits are padded with the minimum similarity of the result set.
unique_chunk_ids, counts = np.unique(chunk_ids, return_counts=True)
score = np.full(
(len(unique_chunk_ids), np.max(counts)), np.min(similarity), dtype=similarity.dtype
)
for i, (unique_chunk_id, count) in enumerate(zip(unique_chunk_ids, counts, strict=True)):
score[i, :count] = similarity[chunk_ids == unique_chunk_id]
pooled_similarity = np.mean(score, axis=1)
# Sort the chunk ids by their adjusted similarity.
sorted_indices = np.argsort(pooled_similarity)[::-1]
unique_chunk_ids = unique_chunk_ids[sorted_indices][:num_results]
pooled_similarity = pooled_similarity[sorted_indices][:num_results]
return unique_chunk_ids.tolist(), pooled_similarity.tolist()
def keyword_search(
query: str, *, num_results: int = 3, config: RAGLiteConfig | None = None
) -> tuple[list[str], list[float]]:
"""Search chunks using BM25 keyword search."""
# Read the config.
config = config or RAGLiteConfig()
db_backend = make_url(config.db_url).get_backend_name()
# Connect to the database.
engine = create_database_engine(config)
with Session(engine) as session:
if db_backend == "postgresql":
# Convert the query to a tsquery [1].
# [1] https://www.postgresql.org/docs/current/textsearch-controls.html
query_escaped = re.sub(r"[&|!():<>\"]", " ", query)
tsv_query = " | ".join(query_escaped.split())
# Perform keyword search with tsvector.
statement = text("""
SELECT id as chunk_id, ts_rank(to_tsvector('simple', body), to_tsquery('simple', :query)) AS score
FROM chunk
WHERE to_tsvector('simple', body) @@ to_tsquery('simple', :query)
ORDER BY score DESC
LIMIT :limit;
""")
results = session.execute(statement, params={"query": tsv_query, "limit": num_results})
elif db_backend == "sqlite":
# Convert the query to an FTS5 query [1].
# [1] https://www.sqlite.org/fts5.html#full_text_query_syntax
query_escaped = re.sub(f"[{re.escape(string.punctuation)}]", "", query)
fts5_query = " OR ".join(query_escaped.split())
# Perform keyword search with FTS5. In FTS5, BM25 scores are negative [1], so we
# negate them to make them positive.
# [1] https://www.sqlite.org/fts5.html#the_bm25_function
statement = text("""
SELECT chunk.id as chunk_id, -bm25(keyword_search_chunk_index) as score
FROM chunk JOIN keyword_search_chunk_index ON chunk.rowid = keyword_search_chunk_index.rowid
WHERE keyword_search_chunk_index MATCH :match
ORDER BY score DESC
LIMIT :limit;
""")
results = session.execute(statement, params={"match": fts5_query, "limit": num_results})
# Unpack the results.
chunk_ids, keyword_score = zip(*results, strict=True)
chunk_ids, keyword_score = list(chunk_ids), list(keyword_score) # type: ignore[assignment]
return chunk_ids, keyword_score # type: ignore[return-value]
def reciprocal_rank_fusion(
rankings: list[list[str]], *, k: int = 60
) -> tuple[list[str], list[float]]:
"""Reciprocal Rank Fusion."""
# Compute the RRF score.
chunk_ids = {chunk_id for ranking in rankings for chunk_id in ranking}
chunk_id_score: defaultdict[str, float] = defaultdict(float)
for ranking in rankings:
chunk_id_index = {chunk_id: i for i, chunk_id in enumerate(ranking)}
for chunk_id in chunk_ids:
chunk_id_score[chunk_id] += 1 / (k + chunk_id_index.get(chunk_id, len(chunk_id_index)))
# Rank RRF results according to descending RRF score.
rrf_chunk_ids, rrf_score = zip(
*sorted(chunk_id_score.items(), key=lambda x: x[1], reverse=True), strict=True
)
return list(rrf_chunk_ids), list(rrf_score)
def hybrid_search(
query: str, *, num_results: int = 3, num_rerank: int = 100, config: RAGLiteConfig | None = None
) -> tuple[list[str], list[float]]:
"""Search chunks by combining ANN vector search with BM25 keyword search."""
# Run both searches.
vs_chunk_ids, _ = vector_search(query, num_results=num_rerank, config=config)
ks_chunk_ids, _ = keyword_search(query, num_results=num_rerank, config=config)
# Combine the results with Reciprocal Rank Fusion (RRF).
chunk_ids, hybrid_score = reciprocal_rank_fusion([vs_chunk_ids, ks_chunk_ids])
chunk_ids, hybrid_score = chunk_ids[:num_results], hybrid_score[:num_results]
return chunk_ids, hybrid_score
def retrieve_chunks(
chunk_ids: list[str],
*,
config: RAGLiteConfig | None = None,
) -> list[Chunk]:
"""Retrieve chunks by their ids."""
config = config or RAGLiteConfig()
engine = create_database_engine(config)
with Session(engine) as session:
chunks = list(session.exec(select(Chunk).where(col(Chunk.id).in_(chunk_ids))).all())
chunks = sorted(chunks, key=lambda chunk: chunk_ids.index(chunk.id))
return chunks
def retrieve_segments(
chunk_ids: list[str] | list[Chunk],
*,
neighbors: tuple[int, ...] | None = (-1, 1),
config: RAGLiteConfig | None = None,
) -> list[str]:
"""Group chunks into contiguous segments and retrieve them."""
# Retrieve the chunks.
config = config or RAGLiteConfig()
chunks: list[Chunk] = (
retrieve_chunks(chunk_ids, config=config) # type: ignore[arg-type,assignment]
if all(isinstance(chunk_id, str) for chunk_id in chunk_ids)
else chunk_ids
)
# Extend the chunks with their neighbouring chunks.
if neighbors:
engine = create_database_engine(config)
with Session(engine) as session:
neighbor_conditions = [
and_(Chunk.document_id == chunk.document_id, Chunk.index == chunk.index + offset)
for chunk in chunks
for offset in neighbors
]
chunks += list(session.exec(select(Chunk).where(or_(*neighbor_conditions))).all())
# Keep only the unique chunks.
chunks = list(set(chunks))
# Sort the chunks by document_id and index (needed for groupby).
chunks = sorted(chunks, key=lambda chunk: (chunk.document_id, chunk.index))
# Group the chunks into contiguous segments.
segments: list[list[Chunk]] = []
for _, group in groupby(chunks, key=lambda chunk: chunk.document_id):
segment: list[Chunk] = []
for chunk in group:
if not segment or chunk.index == segment[-1].index + 1:
segment.append(chunk)
else:
segments.append(segment)
segment = [chunk]
segments.append(segment)
# Rank segments according to the aggregate relevance of their chunks.
chunk_id_to_score = {chunk.id: 1 / (i + 1) for i, chunk in enumerate(chunks)}
segments.sort(
key=lambda segment: sum(chunk_id_to_score.get(chunk.id, 0.0) for chunk in segment),
reverse=True,
)
# Convert the segments into strings.
segments = [
segment[0].headings.strip() + "\n\n" + "".join(chunk.body for chunk in segment).strip() # type: ignore[misc]
for segment in segments
]
return segments # type: ignore[return-value]
def rerank_chunks(
query: str,
chunk_ids: list[str] | list[Chunk],
*,
config: RAGLiteConfig | None = None,
) -> list[Chunk]:
"""Rerank chunks according to their relevance to a given query."""
# Retrieve the chunks.
config = config or RAGLiteConfig()
chunks: list[Chunk] = (
retrieve_chunks(chunk_ids, config=config) # type: ignore[arg-type,assignment]
if all(isinstance(chunk_id, str) for chunk_id in chunk_ids)
else chunk_ids
)
# Early exit if no reranker is configured.
if not config.reranker:
return chunks
# Select the reranker.
if isinstance(config.reranker, Sequence):
# Detect the languages of the chunks and queries.
langs = {detect(str(chunk)) for chunk in chunks}
langs.add(detect(query))
# If all chunks and the query are in the same language, use a language-specific reranker.
rerankers = dict(config.reranker)
if len(langs) == 1 and (lang := next(iter(langs))) in rerankers:
reranker = rerankers[lang]
else:
reranker = rerankers.get("other")
else:
# A specific reranker was configured.
reranker = config.reranker
# Rerank the chunks.
if reranker:
results = reranker.rank(query=query, docs=[str(chunk) for chunk in chunks])
chunks = [chunks[result.doc_id] for result in results.results]
return chunks
|