Use case · Teams building grounded research tools
A cited research assistant
Ground every generated paragraph in real source URLs, because the chunk-to-URL mapping is enforced by the response shape.
The problem
Research assistants that summarise the web are only trustworthy if every claim can be traced to a source. If citations are stitched together in a prompt template after the fact, they drift — the model cites a URL it never actually read. You want attribution guaranteed by the data path, not hoped for at generation time.
How Polymathy helps
- Every chunk the model sees is already paired with the URL it came from, so citations reference text that was genuinely retrieved.
- Sequential u64 chunk IDs give you stable handles for footnote numbering and deduplication.
- Because Polymathy returns the source alongside the text, "which URL did this sentence come from" is always answerable.
What the flow looks like
Research query → Polymathy → chunks with source URLs → LLM synthesis → paragraph with per-sentence citations back to the exact chunk.
A good fit when
- Trustworthy, traceable citations are a hard requirement
- You render footnotes or a sources list from real retrieved text
- You want attribution enforced by the data, not the prompt
Reach for something else when
- You need reranking or answer generation built in (that is downstream of Polymathy)
- Your sources are behind connectors SearxNG cannot reach
Other use cases
A Perplexity-style answer UI
A Rust backend that fetches a handful of pages per query and returns them shaped for prompt-stuffing, so your UI owns the chat and the citations.
Read →An internal-docs answer engine
Point SearxNG at your internal indices and turn matches into cited chunks for a private assistant your employees can trust.
Read →A content-processor test harness
A stable HTTP harness that exercises your content-processor service against real-world URLs drawn from arbitrary queries.
Read →