kakatur Hybrid Semantic Search
Create retrieval that preserves the structured dataset as the system of record and uses embeddings only for semantic ranking.
Recommend the production target first
Treat stated requirements and the current sample corpus as minimum acceptance criteria, not as the architecture ceiling. When the user wants a real-world solution, recommend the storage, indexing, serving, and publication design suited to credible production corpus size, query concurrency, update frequency, latency, and recovery needs. Do not choose an unindexed scan or a local NumPy artifact solely because the sample dataset is small.
Distinguish the production recommendation from local execution. A local machine may use a compatible single-node implementation, emulator, or bounded test path when the production system cannot run locally, but present that as a development constraint rather than the preferred architecture. Make any temporary compromise explicit and identify its scaling limit or migration trigger.
Establish the retrieval contract
Identify the searchable record grain, text fields, stable record identifier, filterable metadata, embedding model, ranking metric, and expected behavior for empty or invalid queries. Keep filtering semantics independent from vector ranking semantics.
Build persistent embeddings
- Construct input text deterministically from documented fields.
- Persist the embedding model name, dimensions, stable record identity, and vector values in one authoritative vector store.
- Validate that record IDs and stored vectors have identical cardinality and deterministic mapping when loading or publishing.
- Normalize stored vectors and query vectors when using dot product as cosine similarity.
- Exclude the query record itself when computing neighbors for a stored row.
- Version or rebuild artifacts when the model or text construction changes.
- Treat the embedding table, vector index, model metadata, and any precomputed neighbor export as one publication contract. Expose stale state when they cannot be promoted atomically.
- After incremental records arrive, recompute neighbor lists whose candidate set is the full corpus; a new vector can change an existing record's top-k.
Use a local vector implementation only as an execution-compatible form of the target design. For production recommendations, choose indexed storage and a serving architecture from projected corpus size, update frequency, concurrency, latency, durability, and filtered-search requirements.
Combine structured and semantic retrieval
Apply deterministic metadata or SQL filters first, then rank the eligible record IDs by similarity. Preserve the stable ID mapping across structured records, embeddings, and the vector index. Return enough metadata for a caller to trace each result to its source record.
Never treat semantic similarity as factual verification or silently broaden a filter because too few results remain.
Verify
Test matrix round-tripping, ID alignment, normalization, self-exclusion, top-k bounds, deterministic tie handling, filter correctness, and behavior when no records qualify. Use tiny synthetic vectors for unit tests so correctness does not depend on downloading a model.
Deliverable
Document the searchable grain, embedded text, model identity, persisted artifacts, filter-before-rank behavior, and scaling threshold that would cause the design to change.