Retrieval
you can check.
Flip a switch on a bucket and every file becomes searchable and citable. Each answer carries a citation you can verify against the exact source — and when an agent uses it, the retrieval lands in the run record.
Bucket → indexed → answered.
BM25 + dense vector retrieval, fused via Reciprocal Rank Fusion (RRF). Top-k 8 hybrid candidates feed the agent. Every answer carries citations back to the source file.
- pricing-faq.md12 KB
- product-overview.pdf482 KB
- release-notes-2026-05.md8 KB
- onboarding-guide.pdf1.2 MB
- top-k8 · hybrid
- encodertext-embedding-3-small
Refunds are processed within 7 business days from the original payment method.
Lexical and semantic, together.
Pure keyword search misses paraphrases. Pure vector search misses exact phrases. We run both legs in parallel and combine them with Reciprocal Rank Fusion — no learned reranker, just the same RRF that Microsoft, Elastic, and the IR community converged on.
Exact terms and phrases — the precise matching you expect from full-text search.
Meaning, not just words. Finds passages that paraphrase the question, even with no shared terms.
Both result lists merged into one ranking — no reranker to train, no extra model call per query.
Every answer carries a receipt.
Show your work.
Every agent answer is paired with a structured receipt — the sources it pulled from, the relevance scores, the chunk hashes, and a verification digest you can check against the index manifest independently. The agent doesn't ask you to trust the answer. It shows you why to.
“What is our refund policy?”
agent · support · 14:02:11 · kr_share_test_3f4d…1c
Refunds are processed within 7 business days from the original payment method.
- pricing-faq.md · §3 · refunds & cancellationsscore · 0.92bucket support-docs · 12 KB · indexed 13:51:30chunk hash 4d2f0e9c7b81a2c9d4e5f6a7…
- billing-policy.md · §1.4 · payment termsscore · 0.74bucket support-docs · 18 KB · indexed 13:51:30chunk hash 4f1ab3a0e7c2f9b8c1d2e3f4…
Anyone — including parties you no longer trust — can independently verify that this answer was assembled from these chunks of these files, at that moment.
Your agent retrieves
with receipts attached.
OpenAI-compatible.
Scoped by default.
Replace the base URL and your existing OpenAI client runs against your bucket. Each agent gets its own scoped credential. Bring your own model key — Kraterion bills $0 for the chat call itself.
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.kraterion.com/v1/agents/support",
apiKey: process.env.KRATERION_KEY,
});
const reply = await client.chat.completions.create({
model: "support",
messages: [
{ role: "user", content: "What is our refund policy?" },
],
});
console.log(reply.choices[0].message.content);
// → Refunds are processed within 7 business days
// from the original payment method.
//
// citations: [pricing-faq.md · §3 · 0.92]What is our refund policy?
- recallquery: "user prefs"→2 notes29 ms
- searchquery: "refund policy"→4 hits62 ms
- readkey: "pricing-faq.md"→12 KB38 ms
Refunds are processed within 7 business days from the original payment method.
Also reachable
from your AI assistant.
Plug Kraterion into Claude Desktop, Cursor, or any MCP-compatible client. Your knowledge, your tools, and persistent memory — over a different protocol, with the same scoped credentials.
{
"mcpServers": {
"kraterion": {
"url": "https://mcp.kraterion.com/mcp",
"auth": {
"type": "oauth",
"dcr": true
}
}
}
}- Claude DesktopFirst-party MCP host
- CursorEditor MCP support
- ZedEditor MCP support
- ContinueIDE assistant
- Any MCP clientImplements the spec
Six tools the agent already knows how to call.
Each one is a typed function with a JSON schema. The agent picks which to call. You can see, log, and override every call — and add your own.
- search
search(bucket, query)Hybrid retrieval over indexed chunks. Returns ranked hits with file + section refs. - list-buckets
list_buckets()Enumerate the buckets the agent's credential is scoped to. - list-objects
list_objects(bucket, prefix?)List S3 keys in a bucket, optionally filtered by prefix. - read
read_object(bucket, key)Fetch full object bytes — decrypted under the agent's scoped credential. - write
write_object(bucket, key, body)Write a new object back to the bucket — only enabled for read-write agents. - manifest
get_manifest(answer_id)Pull the tamper-evident audit record for an earlier agent answer.
Index a bucket.
Get answers with citations.
Citation-bound responses. Drop-in OpenAI clients. Six built-in tools. BYOK for the model — $0 from us on the chat call itself.