feat: Implement hierarchical retrieval architecture (#11610)
This PR implements the complete three-tier hierarchical retrieval architecture
as specified in issue #11610, enabling production-grade RAG capabilities.
## Tier 1: Knowledge Base Routing
- Auto-route queries to relevant knowledge bases
- Per-KB retrieval parameters (KBRetrievalParams dataclass)
- Rule-based routing with keyword overlap scoring
- LLM-based routing with fallback to rule-based
- Configurable routing methods: auto, rule_based, llm_based, all
## Tier 2: Document Filtering
- Document-level metadata filtering within selected KBs
- Configurable metadata fields for filtering
- LLM-generated filter conditions
- Metadata similarity matching (fuzzy matching)
- Enhanced metadata generation for documents
## Tier 3: Chunk Refinement
- Parent-child chunking with summary mapping
- Custom prompts for keyword extraction
- LLM-based question generation for chunks
- Integration with existing retrieval pipeline
## Metadata Management (Batch CRUD)
- MetadataService with batch operations:
- batch_get_metadata
- batch_update_metadata
- batch_delete_metadata_fields
- batch_set_metadata_field
- get_metadata_schema
- search_by_metadata
- get_metadata_statistics
- copy_metadata
- REST API endpoints in metadata_app.py
## Integration
- HierarchicalConfig dataclass for configuration
- Integrated into Dealer class (search.py)
- Wired into agent retrieval tool
- Non-breaking: disabled by default
## Tests
- 48 unit tests covering all components
- Tests for config, routing, filtering, and metadata operations
d104f59e29