Vector Databases and Retrieval Systems

Compare lexical and embedding search on supplied questions.

Evidence: A comparison explains where each method succeeds or fails.

Extract text while retaining headings, source IDs and page references.

Evidence: Parsed chunks retain correct source attribution.

Compare fixed-size and structure-aware chunks.

Evidence: A test set reveals a justified chunk-size choice.

Generate vectors and check dimension, version and normalisation assumptions.

Evidence: Index metadata records the embedding model and dimensions.

Compare exact search and approximate nearest-neighbour concepts.

Evidence: A small benchmark measures recall against exact retrieval.

Restrict retrieval by tenant, document status and permissions.

Evidence: Cross-tenant and unpublished-document tests return no restricted content.

Combine lexical and vector ranking for ambiguous queries.

Evidence: Results improve on at least one declared baseline metric or explain failure.

Rerank candidates and track the added latency.

Evidence: A measured quality and latency tradeoff is documented.

Update changed documents and remove superseded chunks.

Evidence: Deleted content is absent from new queries and audit records remain.

Create relevance labels and calculate retrieval recall and reciprocal rank.

Evidence: Metric calculations match a small hand-checked example.

Investigate empty retrieval, wrong filters and embedding drift.

Evidence: Each injected fault has a reproducible diagnosis.

Expose a search API with timeouts and result provenance.

Evidence: Contract tests verify response structure and graceful failure.