nucliadb 6.4.0.post4200__py3-none-any.whl → 6.4.0.post4210__py3-none-any.whl

nucliadb/search/search/find_merge.py

@@ -42,7 +42,7 @@ from nucliadb.search.search.hydrator import (
 )
 from nucliadb.search.search.merge import merge_relations_results
 from nucliadb.search.search.query_parser.models import UnitRetrieval
-from nucliadb.search.search.rank_fusion import RankFusionAlgorithm
+from nucliadb.search.search.rank_fusion import IndexSource, RankFusionAlgorithm
 from nucliadb.search.search.rerankers import (
     RerankableItem,
     Reranker,
@@ -108,7 +108,13 @@ async def build_find_response(
     )
     graph_results = graph_results_to_text_block_matches(search_response.graph)
 
-    merged_text_blocks = rank_fusion_algorithm.fuse(keyword_results, semantic_results, graph_results)
+    merged_text_blocks = rank_fusion_algorithm.fuse(
+        {
+            IndexSource.KEYWORD: keyword_results,
+            IndexSource.SEMANTIC: semantic_results,
+            IndexSource.GRAPH: graph_results,
+        }
+    )
 
     # cut
     # we assume pagination + predict reranker is forbidden and has been already

nucliadb/search/search/rank_fusion.py

@@ -19,8 +19,10 @@
 #
 import logging
 from abc import ABC, abstractmethod
+from enum import Enum, auto
+from typing import Optional, TypeVar
 
-from nucliadb.common.external_index_providers.base import TextBlockMatch
+from nucliadb.common.external_index_providers.base import ScoredTextBlock
 from nucliadb.common.ids import ParagraphId
 from nucliadb.search.search.query_parser import models as parser_models
 from nucliadb_models.search import SCORE_TYPE
@@ -45,6 +47,14 @@ rank_fusion_observer = Observer(
     ],
 )
 
+ScoredItem = TypeVar("ScoredItem", bound=ScoredTextBlock)
+
+
+class IndexSource(str, Enum):
+    KEYWORD = auto()
+    SEMANTIC = auto()
+    GRAPH = auto()
+
 
 class RankFusionAlgorithm(ABC):
     def __init__(self, window: int):
@@ -60,46 +70,44 @@ class RankFusionAlgorithm(ABC):
         """
         return self._window
 
-    def fuse(
-        self,
-        keyword: list[TextBlockMatch],
-        semantic: list[TextBlockMatch],
-        graph: list[TextBlockMatch],
-    ) -> list[TextBlockMatch]:
-        """Fuse keyword and semantic results and return a list with the merged
+    def fuse(self, sources: dict[str, list[ScoredItem]]) -> list[ScoredItem]:
+        """Fuse elements from multiple sources and return a list of merged
         results.
 
-        If only one retriever is provided, rank fusion will be skipped.
+        If only one source is provided, rank fusion will be skipped.
 
         """
-        # sort results by it's score before merging them
-        keyword = [k for k in sorted(keyword, key=lambda r: r.score, reverse=True)]
-        semantic = [s for s in sorted(semantic, key=lambda r: r.score, reverse=True)]
-        graph = [g for g in graph]
-
-        retrievals_with_results = [x for x in (keyword, semantic, graph) if len(x) > 0]
-        if len(retrievals_with_results) == 1:
-            return retrievals_with_results[0]
+        sources_with_results = [x for x in sources.values() if len(x) > 0]
+        if len(sources_with_results) == 1:
+            # skip rank fusion, we only have a source
+            merged = sources_with_results[0]
         else:
-            merged = self._fuse(keyword, semantic, graph)
+            merged = self._fuse(sources)
+
+        # sort and return the unordered results from the implementation
+        merged.sort(key=lambda r: r.score, reverse=True)
         return merged
 
     @abstractmethod
-    def _fuse(
-        self,
-        keyword: list[TextBlockMatch],
-        semantic: list[TextBlockMatch],
-        graph: list[TextBlockMatch],
-    ) -> list[TextBlockMatch]:
-        """Rank fusion implementation. All arguments are assumed to be ordered
-        by decreasing score."""
+    def _fuse(self, sources: dict[str, list[ScoredItem]]) -> list[ScoredItem]:
+        """Rank fusion implementation.
+
+        Each concrete subclass must provide an implementation that merges
+        `sources`, a group of unordered matches, into a list of unordered
+        results with the new rank fusion score.
+
+        Results can be deduplicated or changed by the rank fusion algorithm.
+
+        """
         ...
 
 
 class ReciprocalRankFusion(RankFusionAlgorithm):
     """Rank-based rank fusion algorithm. Discounts the weight of documents
-    occurring deep in retrieved lists using a reciprocal distribution. It can be
-    parametrized with weights to boost retrievers.
+    occurring deep in retrieved lists using a reciprocal distribution.
+
+    This implementation can be further parametrized with a weight (boost) per
+    retriever that will be applied to all documents ranked by it.
 
     RRF = Σ(r ∈ R) (1 / (k + r(d)) · w(r))
 
@@ -119,9 +127,8 @@ class ReciprocalRankFusion(RankFusionAlgorithm):
         k: float = 60.0,
         *,
         window: int,
-        keyword_weight: float = 1.0,
-        semantic_weight: float = 1.0,
-        graph_weight: float = 1.0,
+        weights: Optional[dict[str, float]] = None,
+        default_weight: float = 1.0,
     ):
         super().__init__(window)
         # Constant used in RRF, studies agree on 60 as a good default value
@@ -129,49 +136,118 @@ class ReciprocalRankFusion(RankFusionAlgorithm):
         # difference among the best results and a smaller score difference among
         # bad results
         self._k = k
-        self._keyword_boost = keyword_weight
-        self._semantic_boost = semantic_weight
-        self._graph_boost = graph_weight
+        self._weights = weights or {}
+        self._default_weight = default_weight
 
     @rank_fusion_observer.wrap({"type": "reciprocal_rank_fusion"})
     def _fuse(
         self,
-        keyword: list[TextBlockMatch],
-        semantic: list[TextBlockMatch],
-        graph: list[TextBlockMatch],
-    ) -> list[TextBlockMatch]:
+        sources: dict[str, list[ScoredItem]],
+    ) -> list[ScoredItem]:
+        # accumulated scores per paragraph
+        scores: dict[ParagraphId, tuple[float, SCORE_TYPE]] = {}
+        # pointers from paragraph to the original source
+        match_positions: dict[ParagraphId, list[tuple[int, int]]] = {}
+
+        # sort results by it's score before fusing them, as we need the rank
+        sources = {
+            retriever: sorted(values, key=lambda r: r.score, reverse=True)
+            for retriever, values in sources.items()
+        }
+        rankings = [
+            (values, self._weights.get(source, self._default_weight))
+            for source, values in sources.items()
+        ]
+        for i, (ranking, weight) in enumerate(rankings):
+            for rank, item in enumerate(ranking):
+                id = item.paragraph_id
+                score, score_type = scores.setdefault(id, (0, item.score_type))
+                score += 1 / (self._k + rank) * weight
+                if {score_type, item.score_type} == {SCORE_TYPE.BM25, SCORE_TYPE.VECTOR}:
+                    score_type = SCORE_TYPE.BOTH
+                scores[id] = (score, score_type)
+
+                position = (i, rank)
+                match_positions.setdefault(item.paragraph_id, []).append(position)
+
+        merged = []
+        for paragraph_id, positions in match_positions.items():
+            # we are getting only one position, effectively deduplicating
+            # multiple matches for the same text block
+            i, j = match_positions[paragraph_id][0]
+            score, score_type = scores[paragraph_id]
+            item = rankings[i][0][j]
+            item.score = score
+            item.score_type = score_type
+            merged.append(item)
+
+        return merged
+
+
+class WeightedCombSum(RankFusionAlgorithm):
+    """Score-based rank fusion algorithm. Multiply each score by a list-specific
+    weight (boost). Then adds the retrieval score of documents contained in more
+    than one list and sort by score.
+
+    wCombSUM = Σ(r ∈ R) (w(r) · S(r, d))
+
+    where:
+    - d is a document
+    - R is the set of retrievers
+    - w(r) weight (boost) for retriever r
+    - S(r, d) is the score of document d given by retriever r
+
+    wCombSUM boosts matches from multiple retrievers and deduplicate them. As a
+    score ranking algorithm, comparison of different scores may lead to bad
+    results.
+
+    """
+
+    def __init__(
+        self,
+        *,
+        window: int,
+        weights: Optional[dict[str, float]] = None,
+        default_weight: float = 1.0,
+    ):
+        super().__init__(window)
+        self._weights = weights or {}
+        self._default_weight = default_weight
+
+    @rank_fusion_observer.wrap({"type": "weighted_comb_sum"})
+    def _fuse(self, sources: dict[str, list[ScoredItem]]) -> list[ScoredItem]:
+        # accumulated scores per paragraph
         scores: dict[ParagraphId, tuple[float, SCORE_TYPE]] = {}
+        # pointers from paragraph to the original source
         match_positions: dict[ParagraphId, list[tuple[int, int]]] = {}
 
         rankings = [
-            (keyword, self._keyword_boost),
-            (semantic, self._semantic_boost),
-            (graph, self._graph_boost),
+            (values, self._weights.get(source, self._default_weight))
+            for source, values in sources.items()
         ]
-        for r, (ranking, boost) in enumerate(rankings):
-            for i, result in enumerate(ranking):
-                id = result.paragraph_id
-                score, score_type = scores.setdefault(id, (0, result.score_type))
-                score += 1 / (self._k + i) * boost
-                if {score_type, result.score_type} == {SCORE_TYPE.BM25, SCORE_TYPE.VECTOR}:
+        for i, (ranking, weight) in enumerate(rankings):
+            for j, item in enumerate(ranking):
+                id = item.paragraph_id
+                score, score_type = scores.setdefault(id, (0, item.score_type))
+                score += item.score * weight
+                if {score_type, item.score_type} == {SCORE_TYPE.BM25, SCORE_TYPE.VECTOR}:
                     score_type = SCORE_TYPE.BOTH
                 scores[id] = (score, score_type)
 
-                position = (r, i)
-                match_positions.setdefault(result.paragraph_id, []).append(position)
+                position = (i, j)
+                match_positions.setdefault(item.paragraph_id, []).append(position)
 
         merged = []
         for paragraph_id, positions in match_positions.items():
             # we are getting only one position, effectively deduplicating
             # multiple matches for the same text block
-            r, i = match_positions[paragraph_id][0]
+            i, j = match_positions[paragraph_id][0]
             score, score_type = scores[paragraph_id]
-            result = rankings[r][0][i]
-            result.score = score
-            result.score_type = score_type
-            merged.append(result)
+            item = rankings[i][0][j]
+            item.score = score
+            item.score_type = score_type
+            merged.append(item)
 
-        merged.sort(key=lambda x: x.score, reverse=True)
         return merged
 
 
@@ -184,9 +260,11 @@ def get_rank_fusion(rank_fusion: parser_models.RankFusion) -> RankFusionAlgorith
         algorithm = ReciprocalRankFusion(
             k=rank_fusion.k,
             window=window,
-            keyword_weight=rank_fusion.boosting.keyword,
-            semantic_weight=rank_fusion.boosting.semantic,
-            graph_weight=rank_fusion.boosting.graph,
+            weights={
+                IndexSource.KEYWORD: rank_fusion.boosting.keyword,
+                IndexSource.SEMANTIC: rank_fusion.boosting.semantic,
+                IndexSource.GRAPH: rank_fusion.boosting.graph,
+            },
         )
 
     else:

nucliadb/writer/api/v1/export_import.py

@@ -25,6 +25,7 @@ from fastapi_versioning import version
 from starlette.requests import Request
 
 from nucliadb.common import datamanagers
+from nucliadb.common.back_pressure import maybe_back_pressure
 from nucliadb.common.cluster.settings import in_standalone_mode
 from nucliadb.common.context import ApplicationContext
 from nucliadb.common.context.fastapi import get_app_context
@@ -45,7 +46,6 @@ from nucliadb.writer import logger
 from nucliadb.writer.api.utils import only_for_onprem
 from nucliadb.writer.api.v1.knowledgebox import create_kb
 from nucliadb.writer.api.v1.router import KB_PREFIX, KBS_PREFIX, api
-from nucliadb.writer.back_pressure import maybe_back_pressure
 from nucliadb_models.export_import import (
     CreateExportResponse,
     CreateImportResponse,
@@ -148,7 +148,7 @@ async def start_kb_import_endpoint(request: Request, kbid: str):
     if not await datamanagers.atomic.kb.exists_kb(kbid=kbid):
         return HTTPClientError(status_code=404, detail="Knowledge Box not found")
 
-    await maybe_back_pressure(request, kbid)
+    await maybe_back_pressure(kbid)
 
     stream = stream_compatible_with_kb(kbid, request.stream())
     try:

nucliadb/writer/api/v1/field.py

@@ -25,6 +25,7 @@ from fastapi_versioning import version
 from starlette.requests import Request
 
 import nucliadb_models as models
+from nucliadb.common.back_pressure import maybe_back_pressure
 from nucliadb.common.maindb.utils import get_driver
 from nucliadb.ingest.orm.knowledgebox import KnowledgeBox
 from nucliadb.models.internal.processing import PushPayload, Source
@@ -40,7 +41,6 @@ from nucliadb.writer.api.v1.resource import (
     validate_rid_exists_or_raise_error,
 )
 from nucliadb.writer.api.v1.router import KB_PREFIX, RESOURCE_PREFIX, RSLUG_PREFIX, api
-from nucliadb.writer.back_pressure import maybe_back_pressure
 from nucliadb.writer.resource.audit import parse_audit
 from nucliadb.writer.resource.field import (
     ResourceClassifications,
@@ -96,7 +96,7 @@ async def add_field_to_resource(
     **parser_kwargs,
 ):
     await validate_rid_exists_or_raise_error(kbid, rid)
-    await maybe_back_pressure(request, kbid, resource_uuid=rid)
+    await maybe_back_pressure(kbid, resource_uuid=rid)
 
     partitioning = get_partitioning()
     partition = partitioning.generate_partition(kbid, rid)
@@ -542,7 +542,7 @@ async def reprocess_file_field(
     x_nucliadb_user: Annotated[str, X_NUCLIADB_USER] = "",
     x_file_password: Annotated[Optional[str], X_FILE_PASSWORD] = None,
 ) -> ResourceUpdated:
-    await maybe_back_pressure(request, kbid, resource_uuid=rid)
+    await maybe_back_pressure(kbid, resource_uuid=rid)
 
     processing = get_processing()
     partitioning = get_partitioning()

nucliadb/writer/api/v1/resource.py

@@ -28,6 +28,7 @@ from fastapi_versioning import version
 from starlette.requests import Request
 
 from nucliadb.common import datamanagers
+from nucliadb.common.back_pressure import maybe_back_pressure
 from nucliadb.common.context.fastapi import get_app_context
 from nucliadb.common.maindb.driver import Driver
 from nucliadb.common.maindb.exceptions import ConflictError, NotFoundError
@@ -45,7 +46,6 @@ from nucliadb.writer.api.v1.router import (
     api,
 )
 from nucliadb.writer.api.v1.slug import ensure_slug_uniqueness, noop_context_manager
-from nucliadb.writer.back_pressure import maybe_back_pressure
 from nucliadb.writer.resource.audit import parse_audit
 from nucliadb.writer.resource.basic import (
     parse_basic_creation,
@@ -106,7 +106,7 @@ async def create_resource(
             detail="Cannot hide a resource: the KB does not have hidden resources enabled",
         )
 
-    await maybe_back_pressure(request, kbid)
+    await maybe_back_pressure(kbid)
 
     partitioning = get_partitioning()
 
@@ -243,7 +243,7 @@ async def modify_resource_endpoint(
 ):
     await validate_rid_exists_or_raise_error(kbid, rid)
 
-    await maybe_back_pressure(request, kbid, resource_uuid=rid)
+    await maybe_back_pressure(kbid, resource_uuid=rid)
 
     if item.slug is None:
         return await modify_resource(
@@ -424,7 +424,7 @@ async def _reprocess_resource(
     x_nucliadb_user: str,
 ):
     await validate_rid_exists_or_raise_error(kbid, rid)
-    await maybe_back_pressure(request, kbid, resource_uuid=rid)
+    await maybe_back_pressure(kbid, resource_uuid=rid)
 
     partitioning = get_partitioning()
 
@@ -571,7 +571,7 @@ async def _reindex_resource(
     reindex_vectors: bool,
 ):
     await validate_rid_exists_or_raise_error(kbid, rid)
-    await maybe_back_pressure(request, kbid, resource_uuid=rid)
+    await maybe_back_pressure(kbid, resource_uuid=rid)
 
     ingest = get_ingest()
     index_req = IndexResource()

nucliadb/writer/api/v1/upload.py

@@ -32,6 +32,7 @@ from fastapi_versioning import version
 from starlette.requests import Request as StarletteRequest
 
 from nucliadb.common import datamanagers
+from nucliadb.common.back_pressure import maybe_back_pressure
 from nucliadb.ingest.orm.utils import set_title
 from nucliadb.models.internal.processing import PushPayload, Source
 from nucliadb.models.responses import HTTPClientError
@@ -43,7 +44,6 @@ from nucliadb.writer.api.v1.resource import (
     validate_rid_exists_or_raise_error,
 )
 from nucliadb.writer.api.v1.slug import ensure_slug_uniqueness, noop_context_manager
-from nucliadb.writer.back_pressure import maybe_back_pressure
 from nucliadb.writer.resource.audit import parse_audit
 from nucliadb.writer.resource.basic import parse_basic_creation, parse_user_classifications
 from nucliadb.writer.resource.field import (
@@ -215,7 +215,7 @@ async def _tus_post(
             detail="Cannot hide a resource: the KB does not have hidden resources enabled",
         )
 
-    await maybe_back_pressure(request, kbid, resource_uuid=path_rid)
+    await maybe_back_pressure(kbid, resource_uuid=path_rid)
 
     dm = get_dm()
     storage_manager = get_storage_manager()
@@ -713,7 +713,7 @@ async def _upload(
     if path_rid is not None:
         await validate_rid_exists_or_raise_error(kbid, path_rid)
 
-    await maybe_back_pressure(request, kbid, resource_uuid=path_rid)
+    await maybe_back_pressure(kbid, resource_uuid=path_rid)
 
     md5_user = x_md5
     path, rid, valid_field = await validate_field_upload(kbid, path_rid, field, md5_user)

nucliadb/writer/lifecycle.py

@@ -21,12 +21,12 @@ from contextlib import asynccontextmanager
 
 from fastapi import FastAPI
 
+from nucliadb.common.back_pressure import start_materializer, stop_materializer
+from nucliadb.common.back_pressure.settings import settings as back_pressure_settings
 from nucliadb.common.context.fastapi import inject_app_context
 from nucliadb.ingest.processing import start_processing_engine, stop_processing_engine
 from nucliadb.ingest.utils import start_ingest, stop_ingest
 from nucliadb.writer import SERVICE_NAME
-from nucliadb.writer.back_pressure import start_materializer, stop_materializer
-from nucliadb.writer.settings import back_pressure_settings
 from nucliadb.writer.tus import finalize as storage_finalize
 from nucliadb.writer.tus import initialize as storage_initialize
 from nucliadb_telemetry.utils import clean_telemetry, setup_telemetry

nucliadb/writer/settings.py

@@ -19,7 +19,6 @@
 #
 from typing import Optional
 
-from pydantic import Field
 from pydantic_settings import BaseSettings
 
 
@@ -29,54 +28,4 @@ class Settings(BaseSettings):
     dm_redis_port: Optional[int] = None
 
 
-class BackPressureSettings(BaseSettings):
-    enabled: bool = Field(
-        default=False,
-        description="Enable or disable back pressure.",
-        alias="back_pressure_enabled",
-    )
-    indexing_rate: float = Field(
-        default=10,
-        description="Estimation of the indexing rate in messages per second. This is used to calculate the try again in time",  # noqa
-    )
-    ingest_rate: float = Field(
-        default=4,
-        description="Estimation of the ingest processed consumer rate in messages per second. This is used to calculate the try again in time",  # noqa
-    )
-    processing_rate: float = Field(
-        default=1,
-        description="Estimation of the processing rate in messages per second. This is used to calculate the try again in time",  # noqa
-    )
-    max_indexing_pending: int = Field(
-        default=1000,
-        description="Max number of messages pending to index in a node queue before rate limiting writes. Set to 0 to disable indexing back pressure checks",  # noqa
-        alias="back_pressure_max_indexing_pending",
-    )
-    max_ingest_pending: int = Field(
-        # Disabled by default
-        default=0,
-        description="Max number of messages pending to be ingested by processed consumers before rate limiting writes. Set to 0 to disable ingest back pressure checks",  # noqa
-        alias="back_pressure_max_ingest_pending",
-    )
-    max_processing_pending: int = Field(
-        default=1000,
-        description="Max number of messages pending to process per Knowledge Box before rate limiting writes. Set to 0 to disable processing back pressure checks",  # noqa
-        alias="back_pressure_max_processing_pending",
-    )
-    indexing_check_interval: int = Field(
-        default=30,
-        description="Interval in seconds to check the indexing pending messages",
-    )
-    ingest_check_interval: int = Field(
-        default=30,
-        description="Interval in seconds to check the ingest pending messages",
-    )
-    max_wait_time: int = Field(
-        default=60,
-        description="Max time in seconds to wait before trying again after back pressure",
-    )
-
-
 settings = Settings()
-
-back_pressure_settings = BackPressureSettings()

{nucliadb-6.4.0.post4200.dist-info → nucliadb-6.4.0.post4210.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nucliadb
-Version: 6.4.0.post4200
+Version: 6.4.0.post4210
 Summary: NucliaDB
 Author-email: Nuclia <nucliadb@nuclia.com>
 License: AGPL
@@ -20,11 +20,11 @@ Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3 :: Only
 Requires-Python: <4,>=3.9
 Description-Content-Type: text/markdown
-Requires-Dist: nucliadb-telemetry[all]>=6.4.0.post4200
-Requires-Dist: nucliadb-utils[cache,fastapi,storages]>=6.4.0.post4200
-Requires-Dist: nucliadb-protos>=6.4.0.post4200
-Requires-Dist: nucliadb-models>=6.4.0.post4200
-Requires-Dist: nidx-protos>=6.4.0.post4200
+Requires-Dist: nucliadb-telemetry[all]>=6.4.0.post4210
+Requires-Dist: nucliadb-utils[cache,fastapi,storages]>=6.4.0.post4210
+Requires-Dist: nucliadb-protos>=6.4.0.post4210
+Requires-Dist: nucliadb-models>=6.4.0.post4210
+Requires-Dist: nidx-protos>=6.4.0.post4210
 Requires-Dist: nucliadb-admin-assets>=1.0.0.post1224
 Requires-Dist: nuclia-models>=0.24.2
 Requires-Dist: uvicorn[standard]

{nucliadb-6.4.0.post4200.dist-info → nucliadb-6.4.0.post4210.dist-info}/RECORD

@@ -60,6 +60,11 @@ nucliadb/common/ids.py,sha256=4QjoIofes_vtKj2HsFWZf8VVIVWXxdkYtLpx1n618Us,8239
 nucliadb/common/locking.py,sha256=RL0CabZVPzxHZyUjYeUyLvsJTm7W3J9o4fEgsY_ufNc,5896
 nucliadb/common/nidx.py,sha256=3EeQGjM_gxK0l_Rb54fspFWVNnzUiKF-_GMxTiiDC8Q,9116
 nucliadb/common/vector_index_config.py,sha256=LqGwhrDCp1q1vBow3scd1Chhr4GLYjYnGL72FKvOYYc,1552
+nucliadb/common/back_pressure/__init__.py,sha256=paAcAZcfGRTyURF9lnn3vX0vcwakTEVswG_xcdGBH-U,928
+nucliadb/common/back_pressure/cache.py,sha256=ANvXglWzI5naAD6N4E_fNi17qS6KNyAhjLeh6WlZZ84,2931
+nucliadb/common/back_pressure/materializer.py,sha256=YzYfN7xI5nlmSowbdLktWIkrJJb3Q2vEmoyz9O3eb2s,11667
+nucliadb/common/back_pressure/settings.py,sha256=3qNOzbI0KC6LMy-wMilXRSBfZu6CCpGHod26MTgAZ2o,3082
+nucliadb/common/back_pressure/utils.py,sha256=aZeP1XSkdgaRgZC76yR9Kje3511ZUCp7KB-XzcvhMYY,2018
 nucliadb/common/cluster/__init__.py,sha256=cp15ZcFnHvpcu_5-aK2A4uUyvuZVV_MJn4bIXMa20ks,835
 nucliadb/common/cluster/exceptions.py,sha256=t7v_l93t44l2tQpdQXgO_w-c4YZRcaayOz1A2i0w4RQ,1258
 nucliadb/common/cluster/grpc_node_dummy.py,sha256=JkufazWzMA4KFEU8EBkMbiiDW4C8lLcRhiiCxP7aCQY,2949
@@ -88,7 +93,7 @@ nucliadb/common/datamanagers/synonyms.py,sha256=zk3GEH38KF5vV_VcuL6DCg-2JwgXJfQl
 nucliadb/common/datamanagers/utils.py,sha256=McHlXvE4P3x-bBY3pr0n8djbTDQvI1G5WusJrnRdhLA,1827
 nucliadb/common/datamanagers/vectorsets.py,sha256=ciYb5uD435Zo8ZbqgPUAszFW9Svp_-R2hY2FEhQ411Y,4304
 nucliadb/common/external_index_providers/__init__.py,sha256=cp15ZcFnHvpcu_5-aK2A4uUyvuZVV_MJn4bIXMa20ks,835
-nucliadb/common/external_index_providers/base.py,sha256=K5lJc2IMcbMTm5rFeB96-mPgicBtGmaunF6GeE2qBxM,8863
+nucliadb/common/external_index_providers/base.py,sha256=BL3DuYbnp-KCmGUiN-FGRtgjWj3SmtgMsGdjGq_7cX4,8905
 nucliadb/common/external_index_providers/exceptions.py,sha256=nDhhOIkb66hjCrBk4Spvl2vN1SuW5gbwrMCDmrdjHHE,1209
 nucliadb/common/external_index_providers/manager.py,sha256=aFSrrKKYG1ydpTSyq4zYD0LOxFS7P-CO6rcKC0hiF4I,4267
 nucliadb/common/external_index_providers/pinecone.py,sha256=sCvMgmXrLebnFLkpLnimDGrm4lE9ydb_ywotOMUSMrI,38124
@@ -234,7 +239,7 @@ nucliadb/search/search/exceptions.py,sha256=klGLgAGGrXcSGix_W6418ZBMqDchAIGjN77o
 nucliadb/search/search/fetch.py,sha256=eiljOKim-4OOEZn-3fyVZSYxztCH156BXYdqlIwVdN4,6181
 nucliadb/search/search/filters.py,sha256=1MkHlJjAQqoRCj7e5cEzK2HvBxGLE17I_omsjiklbtw,6476
 nucliadb/search/search/find.py,sha256=WTc_0HdbaU0Mwkmlf9s4pCmTuU0hz1jetBjIpNLXEEM,7982
-nucliadb/search/search/find_merge.py,sha256=2NtVZlR61MnTZl7mkzqcpIOPUU8hGWPrrP7ATeubt7w,19677
+nucliadb/search/search/find_merge.py,sha256=c-7IlfjfdmWAvQOyM7IO3bKS1EQpnR4oi6pN6mwrQKw,19815
 nucliadb/search/search/graph_merge.py,sha256=y5V7X-BhjHsKDXE69tzQLIIKGm4XuaFrZXw0odcHVNM,3402
 nucliadb/search/search/graph_strategy.py,sha256=JsV-i9PGhekCAzmGpqeueQIitJb7fWCihIwUf76Q3pU,32912
 nucliadb/search/search/hydrator.py,sha256=-R37gCrGxkyaiHQalnTWHNG_FCx11Zucd7qA1vQCxuw,6985
@@ -245,13 +250,13 @@ nucliadb/search/search/paragraphs.py,sha256=pNAEiYqJGGUVcEf7xf-PFMVqz0PX4Qb-WNG-
 nucliadb/search/search/pgcatalog.py,sha256=s_J98fsX_RuFXwpejpkGqG-tD9ELuzz4YQ6U3ew5h2g,9313
 nucliadb/search/search/predict_proxy.py,sha256=IFI3v_ODz2_UU1XZnyaD391fE7-2C0npSmj_HmDvzS4,3123
 nucliadb/search/search/query.py,sha256=-gvKsyGmKYpsoEVzKkq3HJUMcs_3LD3TYUueOcJsTec,11511
-nucliadb/search/search/rank_fusion.py,sha256=8Bq3AHfHgytp7UIhJQgbnXrchrjlCQbgf8PhLQkWBPY,6718
+nucliadb/search/search/rank_fusion.py,sha256=xZtXhbmKb_56gs73u6KkFm2efvTATOSMmpOV2wrAIqE,9613
 nucliadb/search/search/rerankers.py,sha256=PvhExUb8zZYghiFHRgGotw6h6bU--Rft09wE8arvtAw,7424
 nucliadb/search/search/shards.py,sha256=mc5DK-MoCv9AFhlXlOFHbPvetcyNDzTFOJ5rimK8PC8,2636
 nucliadb/search/search/summarize.py,sha256=ksmYPubEQvAQgfPdZHfzB_rR19B2ci4IYZ6jLdHxZo8,4996
 nucliadb/search/search/utils.py,sha256=ajRIXfdTF67dBVahQCXW-rSv6gJpUMPt3QhJrWqArTQ,2175
 nucliadb/search/search/chat/__init__.py,sha256=cp15ZcFnHvpcu_5-aK2A4uUyvuZVV_MJn4bIXMa20ks,835
-nucliadb/search/search/chat/ask.py,sha256=KO1YD1nGwblT4h2CfV7ccCpLP0OuoBCJQHBB5bDQatg,37476
+nucliadb/search/search/chat/ask.py,sha256=jYOGh2rySV4aFx_D2KlNVbPXHBsbkcy0Ve-eBS7CSYc,37611
 nucliadb/search/search/chat/exceptions.py,sha256=Siy4GXW2L7oPhIR86H3WHBhE9lkV4A4YaAszuGGUf54,1356
 nucliadb/search/search/chat/images.py,sha256=PA8VWxT5_HUGfW1ULhKTK46UBsVyINtWWqEM1ulzX1E,3095
 nucliadb/search/search/chat/prompt.py,sha256=Jnja-Ss7skgnnDY8BymVfdeYsFPnIQFL8tEvcRXTKUE,47356
@@ -328,28 +333,27 @@ nucliadb/train/generators/token_classifier.py,sha256=DdyMbrpxIVGWdTcz3SEN_3HwxKf
 nucliadb/train/generators/utils.py,sha256=ZNwvEVPZr-eP0MW3ABN7a11hPQKaa0NdVaRcgBcTp5w,3601
 nucliadb/writer/__init__.py,sha256=S298mrZL3vr62OrBqi97mdLxgR5cReMlRJgnaQHZV7s,1304
 nucliadb/writer/app.py,sha256=ABBO8-u4pDAa61b3mCdD0TFhuHAYcxMkgpZSGgWARuE,2736
-nucliadb/writer/back_pressure.py,sha256=4OwFGq9pvAbChB3WBZAY36lclfD-gD2ouC6YsKA4bIo,16892
 nucliadb/writer/exceptions.py,sha256=-Z7LW--eid7PNeKFuzo9kAlbLEBMUosxE-UVIgGD3SA,929
-nucliadb/writer/lifecycle.py,sha256=OYyhUZ1ejlybPzO-O_EsInjdifKiPiEzooy2d_2DW3k,2550
+nucliadb/writer/lifecycle.py,sha256=P1b_KoNkMTeF1IbyDCh_zhexWbeYe5LH6p2iFSJPiN4,2576
 nucliadb/writer/openapi.py,sha256=thqCO1ht_RJgOkXs-aIsv8aXJrU5z8wo2n05l2_LqMs,1032
 nucliadb/writer/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 nucliadb/writer/run.py,sha256=euVZ_rtHDXs-O1kB-Pt1Id8eft9CYVpWH3zJzEoEqls,1448
-nucliadb/writer/settings.py,sha256=pA9aMAvY8H6zvsxAOdGY8SZLrThDvJ8KLhluGI0GxnQ,3288
+nucliadb/writer/settings.py,sha256=gKtCTDF2E1m6lYL0Iv4WwY4VZuvw1Dsa-uIBZxCHTdU,1071
 nucliadb/writer/utilities.py,sha256=AZ5qEny1Xm0IDsFtH13oJa2usvJZK8f0FdgF1LrnLCw,1036
 nucliadb/writer/api/__init__.py,sha256=cp15ZcFnHvpcu_5-aK2A4uUyvuZVV_MJn4bIXMa20ks,835
 nucliadb/writer/api/constants.py,sha256=qWEDjFUycrEZnSJyLnNK4PQNodU2oVmkO4NycaEZtio,1738
 nucliadb/writer/api/utils.py,sha256=wIQHlU8RQiIGVLI72suvyVIKlCU44Unh0Ae0IiN6Qwo,1313
 nucliadb/writer/api/v1/__init__.py,sha256=akI9A_jloNLb0dU4T5zjfdyvmSAiDeIdjAlzNx74FlU,1128
-nucliadb/writer/api/v1/export_import.py,sha256=elf-EQY5DD3mhw8kWb9tQpDcbrF9sY6VFYqxQOjuVP0,8201
-nucliadb/writer/api/v1/field.py,sha256=KOOBqBJzwsNczn_isxl-YFBL-bmduz3rzSDWMbAJefc,18523
+nucliadb/writer/api/v1/export_import.py,sha256=v0sU55TtRSqDzwkDgcwv2uSaqKCuQTtGcMpYoHQYBQA,8192
+nucliadb/writer/api/v1/field.py,sha256=OicvLF1bnkJj1ixALFLuhvFX6NCMFpORROcFcS9nKpk,18505
 nucliadb/writer/api/v1/knowledgebox.py,sha256=PHEYDFa-sN5JrI8-EiVVg5FDOsRuCLT43kyAB4xt-xA,9530
 nucliadb/writer/api/v1/learning_config.py,sha256=CKBjqcbewkfPwGUPLDWzZSpro6XkmCaVppe5Qtpu5Go,3117
-nucliadb/writer/api/v1/resource.py,sha256=jxphiyeXJq342BR1R8pRQ81L0i3Tczf_Yarqx_DqvWs,19786
+nucliadb/writer/api/v1/resource.py,sha256=IaKHwP4M4Pm3xXj_xcnQCnTzKtXj_xj-r7YOHdH-89I,19750
 nucliadb/writer/api/v1/router.py,sha256=RjuoWLpZer6Kl2BW_wznpNo6XL3BOpdTGqXZCn3QrrQ,1034
 nucliadb/writer/api/v1/services.py,sha256=3AUjk-SmvqJx76v7y89DZx6oyasojPliGYeniRQjpcU,13337
 nucliadb/writer/api/v1/slug.py,sha256=xlVBDBpRi9bNulpBHZwhyftVvulfE0zFm1XZIWl-AKY,2389
 nucliadb/writer/api/v1/transaction.py,sha256=d2Vbgnkk_-FLGSTt3vfldwiJIUf0XoyD0wP1jQNz_DY,2430
-nucliadb/writer/api/v1/upload.py,sha256=fwWXA5BuLPuGKhOcuyf0CdutWJITjJ6fAvDzV_X9VsU,33809
+nucliadb/writer/api/v1/upload.py,sha256=vdKurdxRU7vYlcQIXf5RNTuX-G0waBSak2HnNRmAbLk,33791
 nucliadb/writer/api/v1/vectorsets.py,sha256=F3iMViL5G95_Tns4aO2SOA0DwAzxK2_P8MXxtd_XLRE,6973
 nucliadb/writer/resource/__init__.py,sha256=cp15ZcFnHvpcu_5-aK2A4uUyvuZVV_MJn4bIXMa20ks,835
 nucliadb/writer/resource/audit.py,sha256=FvxMZPzrNHtd31HgpZEvxzwAkbxJTZRhPLqRYYJi3tA,1426
@@ -365,8 +369,8 @@ nucliadb/writer/tus/local.py,sha256=7jYa_w9b-N90jWgN2sQKkNcomqn6JMVBOVeDOVYJHto,
 nucliadb/writer/tus/s3.py,sha256=vF0NkFTXiXhXq3bCVXXVV-ED38ECVoUeeYViP8uMqcU,8357
 nucliadb/writer/tus/storage.py,sha256=ToqwjoYnjI4oIcwzkhha_MPxi-k4Jk3Lt55zRwaC1SM,2903
 nucliadb/writer/tus/utils.py,sha256=MSdVbRsRSZVdkaum69_0wku7X3p5wlZf4nr6E0GMKbw,2556
-nucliadb-6.4.0.post4200.dist-info/METADATA,sha256=mubKUtJdgnEEdrwTcMBVP2xDBYfxCTsqDlxFa3TQugU,4223
-nucliadb-6.4.0.post4200.dist-info/WHEEL,sha256=0CuiUZ_p9E4cD6NyLD6UG80LBXYyiSYZOKDm5lp32xk,91
-nucliadb-6.4.0.post4200.dist-info/entry_points.txt,sha256=XqGfgFDuY3zXQc8ewXM2TRVjTModIq851zOsgrmaXx4,1268
-nucliadb-6.4.0.post4200.dist-info/top_level.txt,sha256=hwYhTVnX7jkQ9gJCkVrbqEG1M4lT2F_iPQND1fCzF80,20
-nucliadb-6.4.0.post4200.dist-info/RECORD,,
+nucliadb-6.4.0.post4210.dist-info/METADATA,sha256=SB9gIMgWxoWNtUEexRLH85E0PL-MnroGhJ6aOambTT4,4223
+nucliadb-6.4.0.post4210.dist-info/WHEEL,sha256=0CuiUZ_p9E4cD6NyLD6UG80LBXYyiSYZOKDm5lp32xk,91
+nucliadb-6.4.0.post4210.dist-info/entry_points.txt,sha256=XqGfgFDuY3zXQc8ewXM2TRVjTModIq851zOsgrmaXx4,1268
+nucliadb-6.4.0.post4210.dist-info/top_level.txt,sha256=hwYhTVnX7jkQ9gJCkVrbqEG1M4lT2F_iPQND1fCzF80,20
+nucliadb-6.4.0.post4210.dist-info/RECORD,,