o ia1@s|UdZddlmZmZmZmZddlZddlZddlm Z e e Z dZ dZGdddZdaeeed <d efd d ZdS) zG Graph-based Reranking Service Advanced reranking using graph features )ListDictAnyOptionalN) Neo4jClientg333333?g333333?c@seZdZdZdefddZ  ddeeee fdede d e d e eee fd eeee ff d d Z dee de d e d ee e ffddZdee de d e d ee e ffddZdee de d e d ee e ffddZdS) GraphRerankerz\ Graph-based reranking service Uses graph features to improve retrieval quality neo4j_clientcCs ||_dS)Nr)selfrr I/lsinfo/ai/hellotax_ai/base_platform/app/services/graph/graph_reranker.py__init__s zGraphReranker.__init__Nresultsquery tenant_idkb_idweightsreturnc Cs|s|S|durddddd}dd|D}i}|durWz ddlm} || j| j| j|} d d | D}WntyV} zt d | WYd} ~ nd} ~ ww| |||} | |||} | |||}|D]z}|d }|d d}| |d}| |d}||d}|d||d||d||d|}||d}|dkr|t9}d|d<d|d<n|dkr|t9}d|d<d|d<n||d<d|d<||d<||d<||d<||d<||d <qn|jdd dd!td"d#|D}t d$t|d%|d&|S)'u Rerank results using graph features Args: results: Initial retrieval results query: Query text tenant_id: Tenant ID kb_id: Knowledge base ID weights: Feature weights (default: balanced) db: SQLAlchemy Session — needed for doc_status penalty lookup. If None, obsolete penalty is skipped. Returns: Reranked results Ng?g?) vector_scorepagerankentity_overlap path_distancecSg|]}|dqS) document_idr .0rr r r Bz(GraphReranker.rerank..r)KnowledgeDocumentcSsi|] }|j|jp dqS) effective)id doc_status)rrowr r r Nz(GraphReranker.rerank..u8[GraphReranker] doc_status 查询失败（非致命）: rscore?rrrrr obsoleter"Tstatus_penalizedamendedForiginal_scorepagerank_scoreentity_overlap_scorepath_distance_scorecSs|dS)Nr&r )xr r r sz&GraphReranker.rerank..)keyreversecss|] }|drdVqdS)r*N)getrr r r sz'GraphReranker.rerank..z Reranked z results using graph features (z obsolete/amended penalized))app.models.knowledge_baserrr!r"filterin_all Exceptionloggerwarning_calculate_pagerank_calculate_entity_overlap_calculate_path_distancesr5_OBSOLETE_SCORE_PENALTY_AMENDED_SCORE_PENALTYsortsuminfolen)r rrrrrdbdoc_idsdoc_status_maprrowsepagerank_scoresentity_overlap_scorespath_distancesresultdoc_idrrrrcombined_scorestatuspenalized_countr r r reranks~              zGraphReranker.rerankrHc sz_d}d}z|jj||||dd}Wn+ty>|jj||||dd}|r1tdd|Dndfdd|D}Ynwd d |D}|r]t|d kr]fd d |D}|WSty}ztd |dd |DWYd}~Sd}~ww)z Calculate PageRank scores for documents Args: doc_ids: Document IDs tenant_id: Tenant ID kb_id: Knowledge base ID Returns: Document ID -> PageRank score mapping aT MATCH (d:Document) WHERE d.id IN $doc_ids AND d.tenant_id = $tenant_id AND d.kb_id = $kb_id CALL gds.pageRank.stream({ nodeProjection: 'Document', relationshipProjection: { REFERENCES: { type: 'REFERENCES', orientation: 'NATURAL' }, SIMILAR_TO: { type: 'SIMILAR_TO', orientation: 'UNDIRECTED' } }, maxIterations: 20, dampingFactor: 0.85 }) YIELD nodeId, score WITH gds.util.asNode(nodeId) AS doc, score WHERE doc.id IN $doc_ids RETURN doc.id as doc_id, score a MATCH (d:Document) WHERE d.id IN $doc_ids AND d.tenant_id = $tenant_id AND d.kb_id = $kb_id OPTIONAL MATCH (d)-[r:REFERENCES|SIMILAR_TO]-(other:Document) WHERE other.tenant_id = $tenant_id AND other.kb_id = $kb_id WITH d.id as doc_id, count(r) as degree RETURN doc_id, degree rHrr parameterscSr)degreer rr r r rrz5GraphReranker._calculate_pagerank..r4cs"g|] }|d|ddqS)rPrXrPr&r r) max_degreer r rscSi|] }|d|dqSrYr rr r r r$r%z5GraphReranker._calculate_pagerank..rci|] \}}||qSr r rkv max_scorer r r$r%zFailed to calculate PageRank: cSi|]}|dqSr(r rrPr r r r$N)r execute_queryr;maxvaluesitemsr<error) r rHrrrfallback_queryrscoresrKr )rZrar r>sF     z!GraphReranker._calculate_pagerankc sz entity overlap score mapping a MATCH (d1:Document) WHERE d1.id IN $doc_ids AND d1.tenant_id = $tenant_id AND d1.kb_id = $kb_id MATCH (d1)-[:CONTAINS]->(e:Entity)<-[:CONTAINS]-(d2:Document) WHERE d2.id IN $doc_ids AND d2.tenant_id = $tenant_id AND d2.kb_id = $kb_id AND d1.id <> d2.id WITH d1.id as doc_id, count(DISTINCT e) as shared_entities RETURN doc_id, shared_entities rUrVcSr[)rPshared_entitiesr rr r r r$r%z;GraphReranker._calculate_entity_overlap..rcr\r r r]r`r r r$r%r(z$Failed to calculate entity overlap: cSrbrcr rdr r r r$#reN)rrfrgrhrir;r<rj) r rHrrrrrlrPrKr r`r r?s0  z'GraphReranker._calculate_entity_overlapc Csz3d}|jj||||dd}i}|D]}|d}t| ||d<q|D] } | |vr0d|| <q&|WStyT} ztd| dd |DWYd } ~ Sd } ~ ww) a Calculate average path distances between documents Args: doc_ids: Document IDs tenant_id: Tenant ID kb_id: Knowledge base ID Returns: Document ID -> path distance score mapping (closer = higher score) a MATCH (d1:Document) WHERE d1.id IN $doc_ids AND d1.tenant_id = $tenant_id AND d1.kb_id = $kb_id MATCH (d2:Document) WHERE d2.id IN $doc_ids AND d2.tenant_id = $tenant_id AND d2.kb_id = $kb_id AND d1.id <> d2.id MATCH path = shortestPath((d1)-[*..3]-(d2)) WITH d1.id as doc_id, avg(length(path)) as avg_distance RETURN doc_id, avg_distance rUrV avg_distancerPg?z$Failed to calculate path distances: cSrb)r'r rdr r r r$_rez;GraphReranker._calculate_path_distances..N)rrfmathexpr;r<rj) r rHrrrrrlrdistancerPrKr r r r@%s. z'GraphReranker._calculate_path_distances)NN)__name__ __module__ __qualname____doc__rr rrstrrintrfloatrTr>r?r@r r r r rsZ  l  a  ; rgraph_rerankerrcCs tdurddlm}t|atS)z%Get or create graph reranker instanceNrr )ryapp.services.graph.neo4j_clientrrr r r r get_graph_rerankerfs r{)rutypingrrrrloggingrorzr getLoggerrrr<rArBrry__annotations__r{r r r r s  Q