Advanced Retrieval Status

Wenlan's released retrieval combines vector embeddings, FTS5 full-text search, graph context, pages, and Reciprocal Rank Fusion. The goal is not a single magic ranking signal; it is a compact context bundle that helps the next AI session continue work.

The public README snapshot reports retrieval metrics, not end-to-end answer quality. Read the evaluation page before using the numbers in comparisons.

Wenlan treats distilled pages as retrieval surfaces, not only as human-readable notes. That matters because repeated project context often becomes clearer as a page than as isolated captures.

A useful recall can include atomic memories, source-backed pages, graph neighbors, and decisions. The best result set is the one that gives the agent enough context without replaying a full chat history.

Recent retrieval work includes a large opt-in batch. Examples include graph gates, temporal filters, query-adaptive routing, FTS hardening, session diversity, salience priors, episode and fact channels, PRF rounds, k-hop graph traversal, global preludes, CoT retrieval, LLM routing, and context compression.

These flags are useful for development and evaluation, but they should not be described as default product behavior until the release notes say so.

A retrieval experiment can improve one benchmark slice and hurt another. Wenlan's eval discipline is deliberately conservative: compare under the same schema, fixture revision, embedder, and run protocol.

When a feature is opt-in, the docs should say which problem it targets and which release status it has instead of implying it is the normal user path.

Most users should care about whether /brief, context, and recall bring the right project facts forward. Maintainers should care about benchmark slices, regression guards, and whether new retrieval work reduces real cold starts.

If recall feels weak, first improve the query and capture quality. Then distill repeated topics into pages before reaching for experimental retrieval flags.