Skip to content

todo.md

Latest commit

 

History

History
55 lines (41 loc) · 2.82 KB

todo.md

File metadata and controls

55 lines (41 loc) · 2.82 KB

1) Fusion Strategies (Vector + Graph) – Principles and Variants

  • Late fusion (recommended baseline): Perform vector top‑k retrieval first, then expand the graph using retrieved entity_ids. This approach is easy to implement, stable, and less prone to semantic drift from the original query.
  • Early fusion: Use the graph to expand the query before vector retrieval. Improves recall but may introduce noise if the entity linking is incorrect.
  • Two‑stage fusion: Conduct vector top‑k retrieval → graph expansion → reranking. Trades latency for improved result quality.
  • Entity‑type gating: Expand the graph only for important entity types (e.g., SECURITY, PROTOCOL, DEVICE, ORG). Reduces noise.
  • Graph expansion depth:
    • Depth 1: Safe (minimal drift), suitable for precise QA.
    • Depth 2+: Enhances recall, suitable for discovery and exploration.
  • Adaptive expansion: Trigger expansion only when score or certainty is high (e.g., top‑1 score above threshold, or entity appears ≥2 times).
  • Score fusion: Combine vector similarity scores with graph relevance metrics (e.g., node centrality, number of relations, path length) for final reranking.
  • Evidence packing: Separate “evidence” (text passages) from “graph facts” (entities/relations) in the constructed prompt.

Strategy selection guidelines:

  • When accuracy is prioritized → Late fusion + Depth 1 + Reranking.
  • When recall and exploration are prioritized → Early fusion + Depth 2 + Entity‑type gating.
  • When latency is prioritized → Late fusion + No reranking + Limited related_k.

2) End‑to‑End Pipeline (Practical Recommendations)

A. Ingestion

  1. Chunk text → generate embeddings → store in Qdrant.
  2. Perform entity and relation extraction → store in Neo4j.
  3. (Optional) Normalize entity names and apply merge policies.

B. Retrieval
4. Conduct vector top‑k search in Qdrant (retrieving passages + entity*ids).
5. *(Optional)_ Filter results by source_id or domain.

C. Graph Expansion
6. Extract entities from entity_ids.
7. Expand relations based on:

  • Entity types (gating).
  • Number of related items (related_k).
  • Expansion depth (start with depth 1, escalate to 2 when necessary).

D. Reranking
8. Define candidates = {passages + graph facts}.
9. Apply reranking using a cross‑encoder or LLM (when available):

  • Prioritize passages containing key entities.
  • Prefer short and high‑confidence relations.

E. Prompt Assembly
10. Evidence block: top passages (concise, with provenance).
11. Graph block: entities + relations (with provenance).
12. Query + Instruction: clearly state reliance on the provided evidence.

F. Answer Generation
13. The LLM generates answers with citations that reference both passages and graph facts.