Retrieval and Ranking

This section explains how the plugin finds relevant chunks and orders them before an answer is generated.

Vector + lexical retrieval (and RRF fusion)

The plugin uses two kinds of search:

• Vector search: finds chunks with similar meaning.
• Lexical search: finds chunks that share important words.

Results from both are combined using RRF (Reciprocal Rank Fusion), which balances semantic and keyword matches.

Retrieval fallback (auto‑broadening)

If the first search looks weak (too few chunks, too little text, or weak scores), the plugin automatically broadens the search and tries again with more candidates.

Optional agentic planner

You can enable agentic retrieval to run a lightweight planner step before answer generation. The planner can keep the current context, retry with expanded retrieval, or pull a full document when whole-document synthesis is needed. Use Agentic max iterations to cap planner loops per query.

Optional query expansion

You can enable query expansion to generate a few alternative queries for short or ambiguous questions. This can improve recall when the exact wording isn’t present in the PDFs.

Optional cross‑encoder reranking

You can enable a cross‑encoder reranker to re‑score the candidate chunks with a more precise model. This is slower but can improve ranking quality.

Tag boosting and max‑per‑doc caps

• Tag boosting can promote chunks whose tags match your query keywords.
• Max‑per‑doc limits how many chunks from the same document appear in the final set, so one PDF doesn’t dominate the answer.

Annotation chunk retrieval

Annotations can be retrieved alongside normal chunks. This lets highlights and notes influence the final answer when they are relevant.