v0.2.0 Rust · actix-web 4 GPL-3.0
Turn search results into an answer-engine chunk map.
Polymathy is a small Rust service that sits in front of a SearxNG instance, pulls the
top result URLs, and hands them to a content processor for chunking and embedding.
It returns chunk_id → (source_url, text) over plain HTTP — the
back of an answer engine, not the front.
$ cargo add polymathy · crates.io/crates/polymathy
GET /v1/search?q=rust+async+patterns
200 OK · application/json
{
"0": ["https://blog.rust-lang.org/…",
"Async functions return a Future…"],
"1": ["https://tokio.rs/…",
"Tokio provides an async runtime…"],
"2": ["https://docs.rs/…",
"When working with async Rust…"]
}What is Polymathy
The fetch-and-chunk seam of an answer engine.
Most teams building answer experiences over web or internal content write the same boring middleware: take a query, get URLs, fetch them, chunk them, return the chunks. Polymathy is exactly that middleware — a single Rust service with an OpenAPI spec — and nothing more.
It is infrastructure for building answer engines, not a turnkey RAG product. You keep ownership of the SearxNG instance, the chunker, the embedding model, and the LLM that turns chunks into a cited paragraph.
WHAT IT ISN'T
- No embedding model bundled — the processor handles that
- No LLM call — Polymathy returns chunks, you cite
- No persistent index — USearch is per-request in v0.2
- No auth or rate limiting — that's your reverse proxy
Problem → solution
Ten blue links
A metasearch engine returns ranked URLs. To answer a question you still have to fetch each page, strip the boilerplate, chunk the text, embed it, and keep track of which chunk came from which URL — per query, in the hot path.
One cited chunk map
Polymathy does the fetch-chunk-embed dance for you and returns a map keyed by
u64 chunk id, each value a (source_url, text) pair. Your
LLM stuffs the text; your UI renders the citation. The source never gets lost.
Features
A narrow surface, precisely shaped.
One endpoint, one job
GET /v1/search?q=… is the whole public surface. No sprawl, no admin console — just the fetch-and-chunk seam an answer engine needs.
Citations by construction
Every chunk carries its source_url in the response shape. The chunk_id → (source_url, text) map makes attribution a data invariant, not a prompt convention.
SearxNG in front
Point at any SearxNG instance via SEARXNG_URL. Polymathy reads its top result URLs — it does not run a crawler or index of its own.
Pluggable processor
Chunking and embedding are delegated to your PROCESSOR_URL service. Swap the chunker or the embedding model without touching Polymathy.
OpenAPI from the types
apistos generates the spec straight from the Rust types. Swagger, ReDoc, RapiDoc, and Scalar UIs ship in the binary.
Boring Rust, single binary
actix-web 4, tokio, reqwest, usearch. One crate, one binary — you supply SearxNG and the processor, nothing else to operate.
Code showcase
The whole service is one request.
GET /v1/search?q={query} hits your SEARXNG_URL, takes the
first ten result URLs, and fans them out to your PROCESSOR_URL in
parallel. The processor returns chunks and 384-dim embeddings; Polymathy assigns
sequential u64 chunk IDs and returns the JSON map.
Everything is generated from Rust types. The OpenAPI spec, Swagger, ReDoc, RapiDoc,
and Scalar UIs come for free via apistos.
# configure the two upstreams
export SEARXNG_URL=https://searx.example.org
export PROCESSOR_URL=https://processor.internal
# run the service
polymathy serve --port 8080
# ask a question
curl 'http://localhost:8080/v1/search?q=how+does+tokio+schedule+tasks'
# -> { "0": ["https://…", "Tokio's scheduler…"], … }How it works
Query in, cited chunk map out.
Four steps, no persistence, no hidden state — the whole request handler on one line.
Honest numbers
No throughput or latency claims — those depend entirely on your SearxNG instance and processor.
Where it fits
Honest fits, not everything.
Perplexity-style UI
A backend that fetches a handful of pages per query and returns them shaped for prompt-stuffing, in Rust.
Read →Internal docs answer engine
Point SearxNG at your internal indices and turn matches into cited chunks for a private assistant.
Read →Chunker test harness
A stable HTTP harness that exercises your content-processor against real-world URLs from arbitrary queries.
Read →Explore the docs
Everything, one hop from here.
About
The contract, the four-step request handler, and an honest account of what Polymathy is and is not.
Read →FAQ
Is it RAG? What it bundles, how persistence works in v0.2, the response shape, and licensing.
Read →Glossary
Answer engine, RAG, chunking, embeddings, SearxNG, metasearch, chunk map — defined as the codebase uses them.
Read →Cited research assistant
Attribution enforced by the response shape, not the prompt — a use case for research-grade answers.
Read →Polymathy vs Onyx
A single fetch-and-chunk service versus the full self-hosted enterprise answer-engine platform.
Read →Polymathy vs Algolia + LLM
Self-hosted, SearxNG-backed retrieval versus hosted SaaS retrieval with per-search pricing.
Read →Read the whole service end to end.
Polymathy is a single GPL-3.0 crate. No lock-in, no SaaS, no hidden index — just the seam between metasearch and your prompt.