Find closer matches,
not just shared keywords.

One body. One short-lived ticket.

You need an existing Claus Network API session and service.call permission on your project. Start with the Network documentation if you haven't set those up.

1. Prepare the exact request bytes

   Serialize a UTF-8 JSON body for /v1/search. This small example uses invented descriptions to illustrate the format, not real projects or measured results.

   {
     "query": "find a tool for checking project links",
     "documents": [
       {"id": "link-tool", "text": "Checks public project links for broken destinations."},
       {"id": "note-tool", "text": "Organizes short research notes by topic."}
     ],
     "limit": 2
   }

   Save that exact byte sequence. Calculate its lowercase hexadecimal SHA-256 hash. Do not reformat or reserialize the body after hashing it.

2. Ask the Network for admission

   Using your Network session, send POST /v1/projects/{yourProjectId}/search-admission to the Network API, with:

   {
     "path": "/v1/search",
     "bodyHash": "<lowercase SHA-256 of the exact UTF-8 body>"
   }

   The resulting ticket is valid for 20 seconds and is bound to this service, this path and this body. Prepare the body first so you can send it promptly.

3. Send the body to the model service

   Send those same bytes to POST https://lhkt280nlpcpgl-8080.proxy.runpod.net/v1/search, with Content-Type: application/json and Authorization: Claus <ticket>.

   Only the short-lived ticket goes to the model service. Never send your Network session credential or a wallet key to that host.

   The response ranks the supplied IDs with cosine-similarity scores from −1 to 1. A score is similarity, not a probability, endorsement or quality rating. Inspect the matched source before relying on it.

Need vectors instead?

Use POST /v1/embeddings with {"input":["first text","second text"]}. Request a ticket bound to that path and the exact body. The response has OpenAI-style data, index, embedding and model fields, with a normalized 384-dimensional vector for each input.

The pretrained Xenova/all-MiniLM-L6-v2 model runs locally on CPU. It is a sentence-embedding model, not a chat model.

Keep requests small

Projects and credentials with the same owner share capacity. Granting service.call lets a collaborator consume that shared allowance. Successful inference consumes capacity even if the client disconnects.

If a request cannot finish

For 429, respect Retry-After. A ticket may already be consumed, so use a fresh ticket for the next bounded attempt rather than replaying it. Keep completed batches and avoid a retry loop. A 503 readiness response is not a successful model check.

The service processes text in memory and does not log request text or retain a shared input cache. This is not permission to submit someone else's private material. Start with public descriptions you are entitled to process.

What was reviewed

This deployment uses the Search-specific owner-operator review policy and Claus's recorded approval of the exact image. It is not independent IMD approval or a comprehensive audit of every third-party dependency.