Search works well on prose but misses exact part numbers. What retrieval change do you recommend?
Why D is the answer
Add keyword search alongside semantic search
Embeddings capture meaning and can miss exact identifiers. Hybrid search handles both.
In detail
- Embeddings capture meaning, and they can miss exact identifiers.
- Keyword search finds exact strings, so hybrid search handles both.
- Test with real queries of both kinds.
Use case
An illustrative scenario, not a real incident.
A parts supplier's search misses "ZX-400". Keyword search finds it, and semantic search still handles "the valve for the small pump", so hybrid search serves both.
See it in different ways
flowchart LR A["Documents"] --> B["Chunk by structure"] B --> C["Index with metadata"] C --> D["Filter and search"] D --> E["Rerank"] E --> F["Answer with citations"]
mindmap
root((Retrieval))
Ingestion
Structure-aware chunks
Search
Keyword
Semantic
Ranking
Reranker
Access
Filter by userflowchart TD
Q{"What is the user searching with?"}
Q -->|"Exact identifier"| A0["Keyword search"]
Q -->|"A description"| A1["Semantic search"]
Q -->|"Either"| A2["Hybrid search"]Key point: Identifiers need keywords, and meaning needs embeddings.
Why the other options are wrong
- A. Switch to a larger embedding model and leave the rest of the search as it is
- A larger embedding model has the same weakness with identifiers.
- B. Ask users to describe the part in a full sentence rather than by its number
- Users know the number, so the system should use it.
- C. Use longer chunks so that each part number has more surrounding context
- More context does not make an exact identifier easier to match.