paper-rag-literature

name	paper-rag-literature
description	Query a configured local PaperRAG corpus for literature review, paper comparison, method tracing, evidence checking, and citation-grounded answers. Use when a request concerns claims, methods, experiments, datasets, limitations, or relationships in PDFs indexed by the local PaperRAG service.

PaperRAG Literature

Use the local PaperRAG service as the evidence authority for questions about the configured named paper libraries.

Workflow

1. Confirm GET http://127.0.0.1:8765/health reports a ready index.
2. Select the intended scope: use paper-rag-all for cross-library synthesis or the matching paper-rag-lib-<id> model for one named library.
3. Retrieve evidence before drafting. Use POST /search with the user's substantive question and optional library_id or source_paths; inspect library names, titles, pages, and excerpts.
4. Answer only from retrieved evidence. Cite every literature claim with the returned source or page identifiers.
5. Label cross-paper or cross-library synthesis as synthesis rather than a direct paper claim.
6. If retrieval is empty or insufficient, state that the selected library scope does not contain enough evidence. Do not substitute model memory.
7. Preserve conditions, datasets, baselines, uncertainty, and negative findings.

For normal end-user dialogue, prefer paper-rag-all unless the user explicitly asks for one library. The server-side policy always repeats retrieval and appends sources regardless of the configured upstream OpenAI-compatible model provider.

For an exact paper scope, use #论文 "filename or title"; for a strict comparison, use #对比 "paper A" | "paper B"; and use #列出论文 to discover names in the currently selected library scope. These directives are resolved locally against indexed metadata. Never guess when PaperRAG reports an ambiguous selector.

For interactive reading, open http://127.0.0.1:8765/workbench. Opening a paper and selecting the answer scope are separate actions: use the checkboxes to select any number of papers and verify their names in the composer scope chips. An empty selection means all indexed papers; a non-empty selection is a hard evidence scope, not a conversational hint. Source links open the indexed PDF but deliberately do not force a page jump because extracted PDF page mappings can be unreliable; use the displayed evidence page number for manual verification.

When translating, send only the text explicitly selected by the user to /api/workbench/translate. Store notes through /api/annotations; annotations are local sidecar data and must never be represented as changes to the original PDF.

Read references/api.md when calling the service programmatically or diagnosing connectivity.