Feat: Medical RAG Pipeline with Multi-Hop Search

[iberi22]

Problem

The MemoryGraph.hybridSearch() supports multi-hop retrieval, but there's no dedicated medical RAG pipeline that:

1. Uses medical context windows (512+ tokens with overlap)
2. Applies re-ranking after initial retrieval
3. Implements query decomposition for complex medical questions
4. Supports evidence citation from retrieved nodes

Proposed Solution

Create lib/src/rag/medical_rag_pipeline.dart that:

1. Extends MemoryPipeline with medical-specific stages:

   • QueryNormalizationStage - expands abbreviations (TA → tensión arterial)
   • MedicalQueryDecompositionStage - splits compound queries
   • HybridRetrievalStage - uses existing hybridSearch with medical config
   • ReRankingStage - applies cross-encoder re-ranking (needs LLM adapter)
   • EvidenceCitationStage - attaches source nodes to generated response

2. Create lib/src/rag/medical_prompt_builder.dart:

   • Constructs RAG prompts with retrieved context
   • Adds medical safety disclaimers
   • Handles Spanish medical terminology

Reference

OrionHealth consumer: orionhealth/lib/features/local_agent/infrastructure/rag_llm_service.dart already uses RagLlmService which needs re-ranking support.

Current limitation noted in IsarVectorStoreService.searchWithReRanking() - the TODO says "Implement re-ranking when LLM adapter is configured"

[iberi22]

@jules-ai crear lib/src/rag/medical_rag_pipeline.dart + medical_prompt_builder.dart. Adjuntar PR pequeno.

[google-labs-jules]

Jules is on it. When finished, you will see another comment and be able to review a PR.

[iberi22]

@jules-ai este issue junto con #41 son críticos. Empezar con #41 (bug fix del dimension hardcoded), luego #38 y #39. El bug blocking los otros features.

[google-labs-jules]

Ready for a review! A PR has been created.

[iberi22]

@jules-ai CORRECCION DEL SCOPE: el issue #39原来的 scope estaba equivocado. isar_agent_memory es un paquete genérico de almacenamiento, NO debe contener lógica de negocio médica.

LO QUE DEBE QUEDAR: solo el framework genérico (MemoryPipeline, hooks de re-ranking interfaces)
LO QUE NO PERTENECE a isar_agent_memory: Medical RAG pipeline, MedicalPromptBuilder, QueryNormalizationStage, MedicalQueryDecompositionStage, EvidenceCitationStage

Estas implementaciones médicas deben vivir en ORIONHEALTH, no en el package core.

Por favor: modifica el PR para que en lugar de crear lib/src/rag/medical_rag_pipeline.dart, crees lib/src/rag/pipeline_hooks.dart con interfaces genéricas:

• ReRankingHook (interface para re-ranking, sin implementación médica)
• QueryExpansionHook (interface, sin lógica de expand)
• CitationHook (interface, sin citation específica)

La implementación médica específica la haré yo en OrionHealth después.