Skip to content

Rerankers

pgvectordb.rerankers

Reranker Module for pgVectorDB

This module provides reranking backends that improve retrieval precision by scoring initial search results against the query using specialized models.

Why Reranking? Bi-encoder models (used for embeddings) are optimized for speed — they encode queries and documents independently. Cross-encoder/reranker models are slower but more accurate: they see both the query and document together, enabling richer comparisons.

Typical workflow
  1. Retrieve top-100 candidates quickly (semantic / hybrid / multimodal search)
  2. Rerank with a cross-encoder or API-based model
  3. Return top-5 or top-10 by rerank score

Supported backends:

Class Backend Requires
CrossEncoderReranker Local sentence-transformers sentence-transformers
CohereReranker Cohere API cohere, API key
AWSBedrockReranker AWS Bedrock boto3, AWS credentials
HuggingFaceReranker Local transformers pipeline transformers, torch

Examples:

>>> from pgvectordb.rerankers import CrossEncoderReranker, CohereReranker
>>>
>>> # Local cross-encoder (no API key needed)
>>> reranker = CrossEncoderReranker(
...     model="cross-encoder/ms-marco-MiniLM-L-6-v2"
... )
>>>
>>> # Rerank via core method
>>> results = await rag.rerank_search(
...     query="best noise cancelling headphones under $200",
...     reranker=reranker,
...     k=100,           # Retrieve 100 candidates
...     rerank_top_k=5,  # Return best 5 after reranking
... )

Version: 0.0.3

Classes

BaseReranker

Bases: ABC

Abstract base class for all reranking backends.

A reranker takes an initial list of search results and re-orders them by computing a relevance score for each (query, document) pair.

Subclass Contract
  • Implement rerank(query, documents, top_k)
  • Return results sorted by rerank score, descending

Examples:

>>> class MyReranker(BaseReranker):
...     def rerank(self, query, documents, top_k=None):
...         # Score each document
...         scored = [(doc, my_score(query, doc)) for doc in documents]
...         scored.sort(key=lambda x: x[1], reverse=True)
...         return [(doc, score) for doc, score in scored[:top_k]]
Source code in pgvectordb\rerankers.py 
 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
class BaseReranker(ABC):
    """
    Abstract base class for all reranking backends.

    A reranker takes an initial list of search results and re-orders them
    by computing a relevance score for each (query, document) pair.

    Subclass Contract:
        - Implement ``rerank(query, documents, top_k)``
        - Return results sorted by rerank score, descending

    Examples:
        >>> class MyReranker(BaseReranker):
        ...     def rerank(self, query, documents, top_k=None):
        ...         # Score each document
        ...         scored = [(doc, my_score(query, doc)) for doc in documents]
        ...         scored.sort(key=lambda x: x[1], reverse=True)
        ...         return [(doc, score) for doc, score in scored[:top_k]]
    """

    @abstractmethod
    def rerank(
        self,
        query: str,
        documents: List[Dict[str, Any]],
        top_k: Optional[int] = None,
    ) -> List[Dict[str, Any]]:
        """
        Rerank a list of documents by relevance to the query.

        Args:
            query: The search query string.
            documents: List of QueryResult-like dicts with at minimum
                ``{"id": ..., "content": ..., "metadata": ..., "score": ...}``.
            top_k: Maximum results to return. If None, returns all input docs
                (reordered).

        Returns:
            Documents re-ordered by rerank score (best first).
            The ``score`` field is replaced with the rerank score.
        """
        ...

    def _safe_top_k(self, documents: List, top_k: Optional[int]) -> int:
        """Return a valid top_k bounded by the number of documents."""
        return min(top_k, len(documents)) if top_k else len(documents)

    def __repr__(self) -> str:
        return f"{self.__class__.__name__}()"
Functions
rerank(query, documents, top_k=None) abstractmethod

Rerank a list of documents by relevance to the query.

Parameters:

Name Type Description Default
query str

The search query string.

 required
documents List[Dict[str, Any]]

List of QueryResult-like dicts with at minimum {"id": ..., "content": ..., "metadata": ..., "score": ...}.

 required
top_k Optional[int]

Maximum results to return. If None, returns all input docs (reordered).

 None

Returns:

Type Description
List[Dict[str, Any]]

Documents re-ordered by rerank score (best first).
List[Dict[str, Any]]

The score field is replaced with the rerank score.
Source code in pgvectordb\rerankers.py 
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
@abstractmethod
def rerank(
    self,
    query: str,
    documents: List[Dict[str, Any]],
    top_k: Optional[int] = None,
) -> List[Dict[str, Any]]:
    """
    Rerank a list of documents by relevance to the query.

    Args:
        query: The search query string.
        documents: List of QueryResult-like dicts with at minimum
            ``{"id": ..., "content": ..., "metadata": ..., "score": ...}``.
        top_k: Maximum results to return. If None, returns all input docs
            (reordered).

    Returns:
        Documents re-ordered by rerank score (best first).
        The ``score`` field is replaced with the rerank score.
    """
    ...

CrossEncoderReranker

Bases: BaseReranker

Local cross-encoder reranker using sentence-transformers.

Cross-encoders process query and document together, providing high accuracy at the cost of latency (no pre-computation possible).

Recommended models
  • cross-encoder/ms-marco-MiniLM-L-6-v2 — Fast, good accuracy (~100ms/batch)
  • cross-encoder/ms-marco-MiniLM-L-12-v2 — Slower, better accuracy
  • cross-encoder/ms-marco-electra-base — Best accuracy, slowest
  • BAAI/bge-reranker-v2-m3 — Multilingual
  • cross-encoder/nli-deberta-v3-small — General NLI (not search-tuned)

Parameters:

Name Type Description Default
model str

HuggingFace model name or local path.

 'cross-encoder/ms-marco-MiniLM-L-6-v2'
device Optional[str]

Device to run on ("cpu", "cuda", "mps"). Auto-detected.

 None
batch_size int

Documents to score per batch (default: 32).

 32
max_length int

Max token length per (query, doc) pair (default: 512).

 512

Examples:

>>> reranker = CrossEncoderReranker(
...     model="cross-encoder/ms-marco-MiniLM-L-6-v2"
... )
>>> results = reranker.rerank(query, candidates, top_k=5)
Source code in pgvectordb\rerankers.py 
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
class CrossEncoderReranker(BaseReranker):
    """
    Local cross-encoder reranker using ``sentence-transformers``.

    Cross-encoders process query and document *together*, providing high
    accuracy at the cost of latency (no pre-computation possible).

    Recommended models:
        - ``cross-encoder/ms-marco-MiniLM-L-6-v2`` — Fast, good accuracy (~100ms/batch)
        - ``cross-encoder/ms-marco-MiniLM-L-12-v2`` — Slower, better accuracy
        - ``cross-encoder/ms-marco-electra-base`` — Best accuracy, slowest
        - ``BAAI/bge-reranker-v2-m3`` — Multilingual
        - ``cross-encoder/nli-deberta-v3-small`` — General NLI (not search-tuned)

    Args:
        model: HuggingFace model name or local path.
        device: Device to run on (``"cpu"``, ``"cuda"``, ``"mps"``). Auto-detected.
        batch_size: Documents to score per batch (default: 32).
        max_length: Max token length per (query, doc) pair (default: 512).

    Examples:
        >>> reranker = CrossEncoderReranker(
        ...     model="cross-encoder/ms-marco-MiniLM-L-6-v2"
        ... )
        >>> results = reranker.rerank(query, candidates, top_k=5)
    """

    def __init__(
        self,
        model: str = "cross-encoder/ms-marco-MiniLM-L-6-v2",
        device: Optional[str] = None,
        batch_size: int = 32,
        max_length: int = 512,
    ):
        try:
            from sentence_transformers import CrossEncoder
        except ImportError:
            raise ImportError(
                "sentence-transformers is required for CrossEncoderReranker. "
                "Install with: pip install sentence-transformers"
            )

        self.model_name = model
        self.batch_size = batch_size
        self.max_length = max_length

        logger.info(f"Loading CrossEncoder: {model}")
        self._model = CrossEncoder(
            model,
            device=device,
            max_length=max_length,
        )
        logger.info(f"✓ CrossEncoder loaded: {model}")

    def rerank(
        self,
        query: str,
        documents: List[Dict[str, Any]],
        top_k: Optional[int] = None,
    ) -> List[Dict[str, Any]]:
        """
        Score each document using the cross-encoder and reorder.

        Args:
            query: Search query.
            documents: Initial search results.
            top_k: Number to return (default: all).

        Returns:
            Reordered documents with updated ``score`` field.
        """
        if not documents:
            return []

        # Build (query, passage) pairs
        pairs = [[query, doc["content"]] for doc in documents]

        # Score all pairs
        scores = self._model.predict(
            pairs,
            batch_size=self.batch_size,
            show_progress_bar=False,
        )

        # Zip and sort
        scored = list(zip(documents, scores.tolist()))
        scored.sort(key=lambda x: x[1], reverse=True)

        top = self._safe_top_k(documents, top_k)
        return [
            {**doc, "score": float(score), "rerank_score": float(score)}
            for doc, score in scored[:top]
        ]

    def __repr__(self) -> str:
        return f"CrossEncoderReranker(model='{self.model_name}')"
Functions
rerank(query, documents, top_k=None)

Score each document using the cross-encoder and reorder.

Parameters:

Name Type Description Default
query str

Search query.

 required
documents List[Dict[str, Any]]

Initial search results.

 required
top_k Optional[int]

Number to return (default: all).

 None

Returns:

Type Description
List[Dict[str, Any]]

Reordered documents with updated score field.
Source code in pgvectordb\rerankers.py 
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
def rerank(
    self,
    query: str,
    documents: List[Dict[str, Any]],
    top_k: Optional[int] = None,
) -> List[Dict[str, Any]]:
    """
    Score each document using the cross-encoder and reorder.

    Args:
        query: Search query.
        documents: Initial search results.
        top_k: Number to return (default: all).

    Returns:
        Reordered documents with updated ``score`` field.
    """
    if not documents:
        return []

    # Build (query, passage) pairs
    pairs = [[query, doc["content"]] for doc in documents]

    # Score all pairs
    scores = self._model.predict(
        pairs,
        batch_size=self.batch_size,
        show_progress_bar=False,
    )

    # Zip and sort
    scored = list(zip(documents, scores.tolist()))
    scored.sort(key=lambda x: x[1], reverse=True)

    top = self._safe_top_k(documents, top_k)
    return [
        {**doc, "score": float(score), "rerank_score": float(score)}
        for doc, score in scored[:top]
    ]

CohereReranker

Bases: BaseReranker

Cloud reranker using the Cohere Rerank API.

Cohere's rerank endpoint is highly optimized for production use — no GPU required locally, and the API handles batching automatically.

Available models
  • rerank-english-v3.0 — English-only, frontier model
  • rerank-multilingual-v3.0 — Multiple languages
  • rerank-english-v2.0 — Legacy English (cheaper)

Pricing: per 1K search units (query + top doc == 1 search unit)

Parameters:

Name Type Description Default
api_key Optional[str]

Cohere API key. If None, reads from COHERE_API_KEY env var.

 None
model str

Rerank model name (default: rerank-english-v3.0).

 'rerank-english-v3.0'
max_chunks_per_doc int

Max text chunks per document (default: 10).

 10

Examples:

>>> import os
>>> reranker = CohereReranker(api_key=os.environ["COHERE_API_KEY"])
>>> results = reranker.rerank(query, candidates, top_k=5)
Source code in pgvectordb\rerankers.py 
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
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
class CohereReranker(BaseReranker):
    """
    Cloud reranker using the Cohere Rerank API.

    Cohere's rerank endpoint is highly optimized for production use — no GPU
    required locally, and the API handles batching automatically.

    Available models:
        - ``rerank-english-v3.0`` — English-only, frontier model
        - ``rerank-multilingual-v3.0`` — Multiple languages
        - ``rerank-english-v2.0`` — Legacy English (cheaper)

    Pricing: per 1K search units (query + top doc == 1 search unit)

    Args:
        api_key: Cohere API key. If None, reads from ``COHERE_API_KEY`` env var.
        model: Rerank model name (default: ``rerank-english-v3.0``).
        max_chunks_per_doc: Max text chunks per document (default: 10).

    Examples:
        >>> import os
        >>> reranker = CohereReranker(api_key=os.environ["COHERE_API_KEY"])
        >>> results = reranker.rerank(query, candidates, top_k=5)
    """

    def __init__(
        self,
        api_key: Optional[str] = None,
        model: str = "rerank-english-v3.0",
        max_chunks_per_doc: int = 10,
    ):
        try:
            import cohere
        except ImportError:
            raise ImportError(
                "cohere is required for CohereReranker. "
                "Install with: pip install cohere"
            )

        import os

        key = api_key or os.environ.get("COHERE_API_KEY")
        if not key:
            raise ValueError(
                "Cohere API key required. Pass api_key= or set COHERE_API_KEY env var."
            )

        self.model = model
        self.max_chunks_per_doc = max_chunks_per_doc
        self._client = cohere.Client(key)
        logger.info(f"CohereReranker initialized: {model}")

    def rerank(
        self,
        query: str,
        documents: List[Dict[str, Any]],
        top_k: Optional[int] = None,
    ) -> List[Dict[str, Any]]:
        """
        Rerank using Cohere Rerank API.

        Args:
            query: Search query.
            documents: Initial search results.
            top_k: Number to return (default: all).

        Returns:
            Reordered documents with updated ``score`` field.
        """
        if not documents:
            return []

        texts = [doc["content"] for doc in documents]
        top = self._safe_top_k(documents, top_k)

        try:
            response = self._client.rerank(
                query=query,
                documents=texts,
                top_n=top,
                model=self.model,
                max_chunks_per_doc=self.max_chunks_per_doc,
            )
        except Exception as e:
            logger.error(f"Cohere rerank API call failed: {e}")
            raise RuntimeError(f"Cohere rerank failed: {e}") from e

        results = []
        for result in response.results:
            doc = documents[result.index]
            score = result.relevance_score
            results.append({**doc, "score": float(score), "rerank_score": float(score)})

        return results

    def __repr__(self) -> str:
        return f"CohereReranker(model='{self.model}')"
Functions
rerank(query, documents, top_k=None)

Rerank using Cohere Rerank API.

Parameters:

Name Type Description Default
query str

Search query.

 required
documents List[Dict[str, Any]]

Initial search results.

 required
top_k Optional[int]

Number to return (default: all).

 None

Returns:

Type Description
List[Dict[str, Any]]

Reordered documents with updated score field.
Source code in pgvectordb\rerankers.py 
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
def rerank(
    self,
    query: str,
    documents: List[Dict[str, Any]],
    top_k: Optional[int] = None,
) -> List[Dict[str, Any]]:
    """
    Rerank using Cohere Rerank API.

    Args:
        query: Search query.
        documents: Initial search results.
        top_k: Number to return (default: all).

    Returns:
        Reordered documents with updated ``score`` field.
    """
    if not documents:
        return []

    texts = [doc["content"] for doc in documents]
    top = self._safe_top_k(documents, top_k)

    try:
        response = self._client.rerank(
            query=query,
            documents=texts,
            top_n=top,
            model=self.model,
            max_chunks_per_doc=self.max_chunks_per_doc,
        )
    except Exception as e:
        logger.error(f"Cohere rerank API call failed: {e}")
        raise RuntimeError(f"Cohere rerank failed: {e}") from e

    results = []
    for result in response.results:
        doc = documents[result.index]
        score = result.relevance_score
        results.append({**doc, "score": float(score), "rerank_score": float(score)})

    return results

AWSBedrockReranker

Bases: BaseReranker

Cloud reranker using AWS Bedrock's rerank endpoint (amazon.rerank-v1:0).

Uses the Amazon Bedrock rerank API, which wraps a hosted reranking model. Requires AWS credentials (via environment, IAM role, or profile).

Available models
  • amazon.rerank-v1:0 — Amazon's primary reranking model
  • cohere.rerank-v3-5:0 — Cohere v3.5 via Bedrock

Parameters:

Name Type Description Default
region_name str

AWS region (default: "us-east-1").

 'us-east-1'
model_id str

Bedrock model ID (default: "amazon.rerank-v1:0").

 'amazon.rerank-v1:0'
aws_access_key_id Optional[str]

Optional, falls back to env/IAM.

 None
aws_secret_access_key Optional[str]

Optional, falls back to env/IAM.

 None
aws_session_token Optional[str]

Optional, for temporary credentials.

 None

Examples:

>>> reranker = AWSBedrockReranker(region_name="us-east-1")
>>> results = reranker.rerank(query, candidates, top_k=5)
Source code in pgvectordb\rerankers.py 
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
class AWSBedrockReranker(BaseReranker):
    """
    Cloud reranker using AWS Bedrock's rerank endpoint (``amazon.rerank-v1:0``).

    Uses the Amazon Bedrock rerank API, which wraps a hosted reranking model.
    Requires AWS credentials (via environment, IAM role, or profile).

    Available models:
        - ``amazon.rerank-v1:0`` — Amazon's primary reranking model
        - ``cohere.rerank-v3-5:0`` — Cohere v3.5 via Bedrock

    Args:
        region_name: AWS region (default: ``"us-east-1"``).
        model_id: Bedrock model ID (default: ``"amazon.rerank-v1:0"``).
        aws_access_key_id: Optional, falls back to env/IAM.
        aws_secret_access_key: Optional, falls back to env/IAM.
        aws_session_token: Optional, for temporary credentials.

    Examples:
        >>> reranker = AWSBedrockReranker(region_name="us-east-1")
        >>> results = reranker.rerank(query, candidates, top_k=5)
    """

    def __init__(
        self,
        region_name: str = "us-east-1",
        model_id: str = "amazon.rerank-v1:0",
        aws_access_key_id: Optional[str] = None,
        aws_secret_access_key: Optional[str] = None,
        aws_session_token: Optional[str] = None,
    ):
        try:
            import boto3
        except ImportError:
            raise ImportError(
                "boto3 is required for AWSBedrockReranker. "
                "Install with: pip install boto3"
            )

        self.model_id = model_id
        self.region_name = region_name

        session_kwargs = {}
        if aws_access_key_id:
            session_kwargs["aws_access_key_id"] = aws_access_key_id
        if aws_secret_access_key:
            session_kwargs["aws_secret_access_key"] = aws_secret_access_key
        if aws_session_token:
            session_kwargs["aws_session_token"] = aws_session_token

        import boto3

        session = boto3.Session(**session_kwargs) if session_kwargs else boto3.Session()
        self._client = session.client("bedrock-agent-runtime", region_name=region_name)
        logger.info(f"AWSBedrockReranker initialized: {model_id} in {region_name}")

    def rerank(
        self,
        query: str,
        documents: List[Dict[str, Any]],
        top_k: Optional[int] = None,
    ) -> List[Dict[str, Any]]:
        """
        Rerank using AWS Bedrock Rerank API.

        Args:
            query: Search query.
            documents: Initial search results.
            top_k: Number to return (default: all).

        Returns:
            Reordered documents with updated ``score`` field.
        """
        if not documents:
            return []

        top = self._safe_top_k(documents, top_k)

        # Build Bedrock sources format
        sources = [
            {
                "inlineDocumentSource": {
                    "textDocument": {"text": doc["content"]},
                    "type": "TEXT",
                }
            }
            for doc in documents
        ]

        try:
            response = self._client.rerank(
                rerankingConfiguration={
                    "bedrockRerankingConfiguration": {
                        "modelConfiguration": {
                            "modelArn": f"arn:aws:bedrock:{self.region_name}::foundation-model/{self.model_id}"
                        },
                        "numberOfResults": top,
                    },
                    "type": "BEDROCK_RERANKING_MODEL",
                },
                sources=sources,
                textSources=[{"text": query}],
            )
        except Exception as e:
            logger.error(f"AWS Bedrock rerank API call failed: {e}")
            raise RuntimeError(f"AWS Bedrock rerank failed: {e}") from e

        results = []
        for item in response.get("rerankingResults", []):
            original_idx = item["index"]
            score = item["relevanceScore"]
            doc = documents[original_idx]
            results.append({**doc, "score": float(score), "rerank_score": float(score)})

        return results

    def __repr__(self) -> str:
        return (
            f"AWSBedrockReranker(model='{self.model_id}', region='{self.region_name}')"
        )
Functions
rerank(query, documents, top_k=None)

Rerank using AWS Bedrock Rerank API.

Parameters:

Name Type Description Default
query str

Search query.

 required
documents List[Dict[str, Any]]

Initial search results.

 required
top_k Optional[int]

Number to return (default: all).

 None

Returns:

Type Description
List[Dict[str, Any]]

Reordered documents with updated score field.
Source code in pgvectordb\rerankers.py 
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
def rerank(
    self,
    query: str,
    documents: List[Dict[str, Any]],
    top_k: Optional[int] = None,
) -> List[Dict[str, Any]]:
    """
    Rerank using AWS Bedrock Rerank API.

    Args:
        query: Search query.
        documents: Initial search results.
        top_k: Number to return (default: all).

    Returns:
        Reordered documents with updated ``score`` field.
    """
    if not documents:
        return []

    top = self._safe_top_k(documents, top_k)

    # Build Bedrock sources format
    sources = [
        {
            "inlineDocumentSource": {
                "textDocument": {"text": doc["content"]},
                "type": "TEXT",
            }
        }
        for doc in documents
    ]

    try:
        response = self._client.rerank(
            rerankingConfiguration={
                "bedrockRerankingConfiguration": {
                    "modelConfiguration": {
                        "modelArn": f"arn:aws:bedrock:{self.region_name}::foundation-model/{self.model_id}"
                    },
                    "numberOfResults": top,
                },
                "type": "BEDROCK_RERANKING_MODEL",
            },
            sources=sources,
            textSources=[{"text": query}],
        )
    except Exception as e:
        logger.error(f"AWS Bedrock rerank API call failed: {e}")
        raise RuntimeError(f"AWS Bedrock rerank failed: {e}") from e

    results = []
    for item in response.get("rerankingResults", []):
        original_idx = item["index"]
        score = item["relevanceScore"]
        doc = documents[original_idx]
        results.append({**doc, "score": float(score), "rerank_score": float(score)})

    return results

HuggingFaceReranker

Bases: BaseReranker

Local reranker using HuggingFace transformers text-classification pipeline.

Any sequence-classification model that scores (query, document) pairs can be used here. This is the most flexible option for custom or fine-tuned models.

Recommended models
  • BAAI/bge-reranker-v2-m3 — Strong multilingual reranker
  • BAAI/bge-reranker-base — Lighter version
  • cross-encoder/ms-marco-MiniLM-L-6-v2 — Popular, fast
  • jinaai/jina-reranker-v2-base-multilingual — Jina's multilingual model

Parameters:

Name Type Description Default
model str

HuggingFace model name or local path.

 'BAAI/bge-reranker-v2-m3'
device Optional[str]

Device ("cpu", "cuda", "mps"). Auto-detected if None.

 None
batch_size int

Batch size for inference (default: 16).

 16
max_length int

Max token length (default: 512).

 512

Examples:

>>> reranker = HuggingFaceReranker(
...     model="BAAI/bge-reranker-v2-m3",
...     device="cuda",
... )
>>> results = reranker.rerank(query, candidates, top_k=5)
Source code in pgvectordb\rerankers.py 
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
class HuggingFaceReranker(BaseReranker):
    """
    Local reranker using HuggingFace ``transformers`` text-classification pipeline.

    Any sequence-classification model that scores (query, document) pairs can be
    used here. This is the most flexible option for custom or fine-tuned models.

    Recommended models:
        - ``BAAI/bge-reranker-v2-m3`` — Strong multilingual reranker
        - ``BAAI/bge-reranker-base`` — Lighter version
        - ``cross-encoder/ms-marco-MiniLM-L-6-v2`` — Popular, fast
        - ``jinaai/jina-reranker-v2-base-multilingual`` — Jina's multilingual model

    Args:
        model: HuggingFace model name or local path.
        device: Device (``"cpu"``, ``"cuda"``, ``"mps"``). Auto-detected if None.
        batch_size: Batch size for inference (default: 16).
        max_length: Max token length (default: 512).

    Examples:
        >>> reranker = HuggingFaceReranker(
        ...     model="BAAI/bge-reranker-v2-m3",
        ...     device="cuda",
        ... )
        >>> results = reranker.rerank(query, candidates, top_k=5)
    """

    def __init__(
        self,
        model: str = "BAAI/bge-reranker-v2-m3",
        device: Optional[str] = None,
        batch_size: int = 16,
        max_length: int = 512,
    ):
        try:
            import torch
            from transformers import AutoModelForSequenceClassification, AutoTokenizer
        except ImportError:
            raise ImportError(
                "transformers and torch are required for HuggingFaceReranker. "
                "Install with: pip install transformers torch"
            )

        import torch
        from transformers import AutoModelForSequenceClassification, AutoTokenizer

        self.model_name = model
        self.batch_size = batch_size
        self.max_length = max_length

        # Auto-detect device
        if device is None:
            if torch.cuda.is_available():
                device = "cuda"
            elif torch.backends.mps.is_available():
                device = "mps"
            else:
                device = "cpu"
        self.device = device

        logger.info(f"Loading HuggingFace reranker: {model} on {device}")
        self._tokenizer = AutoTokenizer.from_pretrained(model)
        self._model = AutoModelForSequenceClassification.from_pretrained(model)
        self._model.to(device)
        self._model.eval()
        logger.info(f"✓ HuggingFace reranker loaded: {model}")

    def _score_batch(self, pairs: List[Tuple[str, str]]) -> List[float]:
        """Score a batch of (query, document) pairs."""
        import torch

        inputs = self._tokenizer(
            [q for q, _ in pairs],
            [d for _, d in pairs],
            padding=True,
            truncation=True,
            max_length=self.max_length,
            return_tensors="pt",
        ).to(self.device)

        with torch.no_grad():
            logits = self._model(**inputs).logits

        # For binary classifiers, use logit of class 1 (relevant)
        # For single-output models, use the raw logit
        if logits.shape[-1] == 2:
            scores = torch.softmax(logits, dim=-1)[:, 1]
        else:
            scores = logits.squeeze(-1)

        return scores.cpu().tolist()

    def rerank(
        self,
        query: str,
        documents: List[Dict[str, Any]],
        top_k: Optional[int] = None,
    ) -> List[Dict[str, Any]]:
        """
        Rerank documents using the local HuggingFace model.

        Args:
            query: Search query.
            documents: Initial search results.
            top_k: Number to return (default: all).

        Returns:
            Reordered documents with updated ``score`` field.
        """
        if not documents:
            return []

        all_pairs = [(query, doc["content"]) for doc in documents]
        all_scores = []

        # Process in batches
        for i in range(0, len(all_pairs), self.batch_size):
            batch = all_pairs[i : i + self.batch_size]
            batch_scores = self._score_batch(batch)
            all_scores.extend(batch_scores)

        # Sort by score
        scored = list(zip(documents, all_scores))
        scored.sort(key=lambda x: x[1], reverse=True)

        top = self._safe_top_k(documents, top_k)
        return [
            {**doc, "score": float(score), "rerank_score": float(score)}
            for doc, score in scored[:top]
        ]

    def __repr__(self) -> str:
        return f"HuggingFaceReranker(model='{self.model_name}', device='{self.device}')"
Functions
rerank(query, documents, top_k=None)

Rerank documents using the local HuggingFace model.

Parameters:

Name Type Description Default
query str

Search query.

 required
documents List[Dict[str, Any]]

Initial search results.

 required
top_k Optional[int]

Number to return (default: all).

 None

Returns:

Type Description
List[Dict[str, Any]]

Reordered documents with updated score field.
Source code in pgvectordb\rerankers.py 
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
def rerank(
    self,
    query: str,
    documents: List[Dict[str, Any]],
    top_k: Optional[int] = None,
) -> List[Dict[str, Any]]:
    """
    Rerank documents using the local HuggingFace model.

    Args:
        query: Search query.
        documents: Initial search results.
        top_k: Number to return (default: all).

    Returns:
        Reordered documents with updated ``score`` field.
    """
    if not documents:
        return []

    all_pairs = [(query, doc["content"]) for doc in documents]
    all_scores = []

    # Process in batches
    for i in range(0, len(all_pairs), self.batch_size):
        batch = all_pairs[i : i + self.batch_size]
        batch_scores = self._score_batch(batch)
        all_scores.extend(batch_scores)

    # Sort by score
    scored = list(zip(documents, all_scores))
    scored.sort(key=lambda x: x[1], reverse=True)

    top = self._safe_top_k(documents, top_k)
    return [
        {**doc, "score": float(score), "rerank_score": float(score)}
        for doc, score in scored[:top]
    ]

Functions

create_reranker(backend, **kwargs)

Factory function to create a reranker by backend name.

Parameters:

Name Type Description Default
backend str

One of "cross_encoder", "cohere", "bedrock", "huggingface".

 required
**kwargs Any

Arguments forwarded to the reranker constructor.

 {}

Returns:

Type Description
BaseReranker

Configured reranker instance.

Raises:

Type Description
ValueError

If backend name is unknown.

Examples:

>>> reranker = create_reranker("cross_encoder",
...     model="cross-encoder/ms-marco-MiniLM-L-6-v2"
... )
>>> reranker = create_reranker("cohere", api_key="...", model="rerank-english-v3.0")
>>> reranker = create_reranker("bedrock", region_name="us-west-2")
>>> reranker = create_reranker("huggingface", model="BAAI/bge-reranker-v2-m3")
Source code in pgvectordb\rerankers.py 
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
def create_reranker(
    backend: str,
    **kwargs: Any,
) -> BaseReranker:
    """
    Factory function to create a reranker by backend name.

    Args:
        backend: One of ``"cross_encoder"``, ``"cohere"``, ``"bedrock"``,
            ``"huggingface"``.
        **kwargs: Arguments forwarded to the reranker constructor.

    Returns:
        Configured reranker instance.

    Raises:
        ValueError: If backend name is unknown.

    Examples:
        >>> reranker = create_reranker("cross_encoder",
        ...     model="cross-encoder/ms-marco-MiniLM-L-6-v2"
        ... )
        >>> reranker = create_reranker("cohere", api_key="...", model="rerank-english-v3.0")
        >>> reranker = create_reranker("bedrock", region_name="us-west-2")
        >>> reranker = create_reranker("huggingface", model="BAAI/bge-reranker-v2-m3")
    """
    backends = {
        "cross_encoder": CrossEncoderReranker,
        "crossencoder": CrossEncoderReranker,
        "cohere": CohereReranker,
        "bedrock": AWSBedrockReranker,
        "aws": AWSBedrockReranker,
        "aws_bedrock": AWSBedrockReranker,
        "huggingface": HuggingFaceReranker,
        "hf": HuggingFaceReranker,
    }

    key = backend.lower().replace("-", "_")
    cls = backends.get(key)

    if cls is None:
        raise ValueError(
            f"Unknown reranker backend: '{backend}'. Supported: {list(backends.keys())}"
        )

    return cls(**kwargs)