Glossary
Terms, as Polymathy uses them.
Answer-engine vocabulary is overloaded. These are the definitions that apply on this site and in the codebase — precise about what Polymathy owns and what it delegates.
- Answer engine
- A system that responds to a query with a synthesised, cited answer rather than a list of links. Polymathy provides the retrieval-and-chunking layer of one; the synthesis is a downstream LLM call.
- RAG (retrieval-augmented generation)
- A pattern where a language model’s prompt is augmented with retrieved documents. Polymathy handles the retrieval-and-chunking half; the generation half is yours. Polymathy is not itself a RAG product.
- Chunking
- Splitting a fetched page into smaller passages sized for embedding and prompt-stuffing. Polymathy delegates chunking to the content processor; the default request asks for 100-word, word-based chunks.
- Embedding
- A vector representation of a chunk’s meaning, used for similarity search. The content processor produces embeddings; the default request asks for AllMiniLML6V2, which yields 384-dimensional vectors.
- SearxNG
- An open-source metasearch engine that aggregates results from many upstream search engines. Polymathy reads the top result URLs from a SearxNG instance you configure via SEARXNG_URL.
- Metasearch
- Search that queries multiple engines and merges their results, rather than maintaining its own crawl and index. SearxNG is Polymathy’s metasearch frontend.
- Content processor
- The HTTP service behind PROCESSOR_URL that fetches a URL, strips boilerplate, chunks the text, and embeds each chunk. Polymathy calls it once per result URL and does not implement it.
- Chunk map
- Polymathy’s response: a JSON object mapping sequential u64 chunk IDs to (source_url, text) pairs. It is the whole output of /v1/search.
- Citation by construction
- The property that every chunk arrives paired with its source URL in the response shape, making attribution a data invariant rather than a prompt-time convention.
- USearch
- A compact vector-search library. Polymathy instantiates a 384-dim, inner-product, F32-quantised USearch index per request in src/index.rs; in v0.2 it is wired in but unused on the read path.
- apistos
- A Rust crate that generates an OpenAPI specification directly from actix-web handler types, so Polymathy’s spec cannot drift from its code.
- Fan-out
- Dispatching the top result URLs to the content processor concurrently, so per-query latency tracks the slowest page rather than the sum of all pages.