Retrieval misses documents that contain exact product codes, though semantic matches work well. What do you try?
Why B is the answer
Hybrid search that adds keyword matching
Exact identifiers are a known weakness of embeddings. Hybrid search adds keyword matching and keeps the semantic results.
In detail
- Keyword matching finds exact product codes.
- Hybrid search keeps the semantic matches too.
- Test with real product codes.
Use case
An illustrative scenario, not a real incident.
Searches for "XR-200" miss, although descriptions work. Adding keyword matching through hybrid search finds exact codes as well.
See it in different ways
flowchart TD
A["Poor answers"] --> B{"Right chunks retrieved?"}
B -->|No| C["Tune chunking, embeddings and search"]
B -->|Yes| D["Tune the prompt or model"]stateDiagram-v2 [*] --> CheckRetrieval CheckRetrieval --> TuneChunks: wrong chunks CheckRetrieval --> TunePrompt: right chunks, poor answers TunePrompt --> FineTune: style still not met TuneChunks --> [*] FineTune --> [*]
flowchart TD
Q{"What is missed?"}
Q -->|"Exact codes"| A0["Add keyword search"]
Q -->|"Meaning"| A1["Improve embeddings"]Key point: Exact codes need keyword matching.
Why the other options are wrong
- A. More replicas of the retrieval endpoint
- Replicas add capacity, not better matches.
- C. A larger model for the generation step
- The generator cannot use documents it never received.
- D. A higher similarity threshold on vector search
- A higher threshold returns even fewer results.