We shaped hybrid retrieval around vector and keyword signals. The intent is practical: semantic similarity alone can miss exact terminology, while keyword matching alone can miss paraphrase and context.
The orchestration layer keeps retrieval results and diagnostics separate. That matters because evidence quality needs to be inspected before answer synthesis, not explained after the fact.
Why hybrid
Vector retrieval is useful when meaning matters more than wording. Keyword retrieval is useful when exact identifiers, names, policy language, or error strings matter. Agentic RAG needs both paths available without treating either as universally correct.
Public-safe shape
| Retrieval signal | Useful for | Failure mode |
|---|---|---|
| Vector | Paraphrase, conceptual match, broad semantic context | Can miss exact terms |
| Keyword | Exact phrases, identifiers, commands, named artifacts | Can miss meaning |
| Fusion | Combining complementary signals | Can add noise if not inspected |
Constraint
Hybrid retrieval is not a claim of better accuracy by itself. It only gives the runtime more explicit evidence paths. Accuracy still requires evaluation against ground truth.