Agentic RAG that reads the footnotes first

What it does PaperQA2 is an agentic RAG system built for scientific literature. You point it at a folder of PDFs, text files, or Microsoft Office documents, and it indexes them into a local full-text search engine. When you ask a question, an LLM agent generates keyword queries, retrieves relevant chunks, re-ranks them with contextual summaries, and synthesizes an answer complete with in-text citations and page references. It also automatically fetches metadata like citation counts, journal quality, and retraction status from providers such as Semantic Scholar and Crossref.

The interesting bit The system treats citation as a first-class constraint, not an afterthought. Its three-phase algorithm—paper search, evidence gathering with LLM-based re-scoring, and answer generation—forces the model to justify every claim with specific document excerpts. The December 2025 update even added multimodal support, passing tables, figures, and math equations through the same summarization pipeline so they can be cited like text.

Key highlights

• Agentic workflow that iteratively refines queries and re-ranks evidence before generating an answer.
• Embeds document metadata into vector search and enriches multimodal objects with an enrichment_llm prompt.
• Bundled “known-good” hyperparameter presets and LiteLLM integration for swapping models or running locally.
• Automatic metadata fetching includes citation counts, retraction checks, and journal quality data.
• Switched to calendar versioning in December 2025, explicitly abandoning backwards-compatibility guarantees to avoid “SemVer confusion.”

Caveats

• The authors admit they are “challenged at naming and counting” and reserve the right to rename the project arbitrarily.
• Calendar versioning means no backwards compatibility across releases; pickled Docs objects from prior major versions are already incompatible.
• The FAQ includes a section titled “How come I get different results than your papers?”, suggesting reproducibility can vary with model or settings.

Verdict Researchers and developers building literature-review tools or domain-specific Q&A over document collections should look here. If you just need generic chat over a few PDFs, the complexity and breaking changes are probably overkill.