Hybrid Search: Combine Lexical and Vector Rankings
Combine lexical and vector candidates with deliberate ranking. Explore reciprocal rank fusion, candidate windows, duplication, and segment-level tests.
VECTORTOKEN LAB / TAG
RAG articles
Trace the path from retrieved passage to generated answer.
Retrieval-augmented generation adds source evidence to a model request. It creates several separate questions: whether the correct passage was indexed, whether retrieval found it, whether context selection kept it, and whether the generated answer accurately used it. This collection keeps those questions visible.
Start with the vector LLM architecture article, then use the search guide to evaluate evidence coverage. The hybrid ranking article is useful when exact identifiers and paraphrases require complementary candidate signals. Keep the retrieval task and eligible corpus fixed while testing changes to ranking or generation so improvements can be attributed to a particular stage.
Inspect unsupported requests, conflicting revisions, and malicious instructions embedded in source material. Provide a deliberate no-answer path when evidence is insufficient. Source identifiers should resolve to the passages actually used, and their access checks must remain intact when previews or citations are displayed. Evidence should remain inspectable beyond the final fluent response.
THE READING PATH
03 articles to explore.
Vector LLM Architecture: Build RAG Around Evidence
Design retrieval-augmented generation around traceable evidence, controlled context, access boundaries, and separate retrieval and answer evaluations.
Token Vector Search: A Practical Retrieval Guide
Build an inspectable semantic retrieval baseline. Evaluate compatible embeddings, exact search, relevance labels, permissions, and candidate coverage.