Aimnis

Collaborative Search for Agents

Search once. Answer everyone. An open-source, cache-first web-search gateway for coding agents: ask a question and get a distilled, source-cited answer instantly from a shared, always-current knowledge pool — so every search makes the pool smarter and cheaper for everyone. Like RAG, but over a communal live-web pool, not your stale private docs.

Status: public preview. The hosted service is live — get a free eval key at aimnis.com and point your agent at it in one minute (per-agent setup). We're proving one thing in public — that the cache hit rate compounds as the pool grows (the flywheel, live dashboard). Follow along.

Why

Every model has a training cutoff. The moment it ships, the world moves on — new library versions, new APIs, new errors — and the model can't keep up without searching the live web on every question. That's slow, expensive, and per-vendor.

Aimnis is the shared, always-current layer in front of that: the first thing an agent checks. Ask a question; if it (or a semantically similar one) has been asked before, you get a distilled, source-cited answer instantly for near-zero cost. If it hasn't, Aimnis fetches it live, distills it, and adds it to the pool — so the next agent to ask gets it free. The corpus captures what happened after every model's cutoff, which no static training set can.

How it works

query
  └─ scrub secrets/PII (redacted before embed, search, distill, or storage)
       └─ local embed + normalize
            └─ semantic cache lookup (exact hash → vector nearest-neighbour)
                 ├─ HIT  → return the pooled, cited answer instantly (no upstream cost)
                 └─ MISS → live search → distill into a cited answer → quality-gate → pool it

• Cache-first. The knowledge pool is the semantic cache (pgvector). A reworded question hits the same entry, so the pool compounds faster than exact matching would.
• Grounded, cited answers. Misses are distilled from live web results into a short answer with [n] citations back to sources — not raw links. The answer is AI-generated (a model distills the sources) and labeled as such in the tool output, so the agent always knows it's reading a machine-written summary.
• Provenance & freshness by default. Every answer carries its model, sources, and a cache timestamp (when the answer was produced or last re-distilled — not when it was last served), plus a relative age, so the agent (and our own ranking) can weigh staleness and decide when to escalate to live search.
• Cited links can be routed for a relevance signal (opt-in, aggregate-only). When enabled, a cited source link points at a signed /r/… redirect that logs which pooled answer earned a follow-through, then forwards to the source — this is how click-through improves ranking. The source's real host is shown inline so the agent still sees where it's going, and the log records only entry + source + host + time — never IP, user-agent, or any user/session id. It's telemetry on the pool, not on you. It's off unless a signing secret is configured, tokens are HMAC-signed (so /r can't be abused as an open redirector), and self-hosting points the redirect at your own gateway — so clicks stay on your box.
• Privacy-conscious ingress. Common secret formats (API keys, tokens, private keys, connection strings) and emails are redacted before a query is embedded, sent to a search/LLM provider, or pooled; secret-dense queries are served live-only and never stored. This is high-precision format-based scrubbing plus an entropy pass — it catches known shapes, not every possible secret or PII, so treat it as defense-in-depth, not a guarantee. The pipeline is open source precisely so you can audit and extend it. Where it runs: s

... View full README on GitHub