Knowledge Base and Feedback APIs

The Runtime exposes two endpoints intended for backend integrations:

Both use the Runtime Base URL and a Bearer API key. Call them from a trusted backend; never embed the key in a browser or mobile application.

Shared authentication

Authorization: Bearer sk-your-api-key
Content-Type: application/json

An API key does not grant additional asset permissions. Knowledge base search still checks whether the represented user or workspace can access the requested knowledge base. Feedback must resolve to an accessible Customize Model before it can create memory.

Search a knowledge base

curl -X POST "$UNIEAI_RUNTIME_BASE_URL/v1/kb/search" \
  -H "Authorization: Bearer $UNIEAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"kb_id":"kb_xxx","query":"How many days of paternity leave are available?","mode":"hybrid","limit":8,"min_score":0.18}'

Request fields

Successful response:

{
  "request_id": "req_abc123",
  "data": {
    "kb_id": "kb_xxx",
    "query": "How many days of paternity leave are available?",
    "limit": 8,
    "min_score": 0.18,
    "count": 1,
    "mode": "hybrid",
    "warnings": [],
    "results": [{
      "chunkId": "chunk_xxx",
      "documentId": "doc_xxx",
      "content": "Employees may request paternity leave...",
      "score": 0.82,
      "pageIndex": 2,
      "pageImageUrl": null,
      "figureImageUrl": null,
      "metadata": {}
    }]
  }
}

Result fields vary by chunk type. For page or Figure assets, the Runtime materializes externally loadable image URLs and replaces inline knowledge base asset keys in content. These URLs can expire and should not be stored permanently.

Submit feedback

curl -X POST "$UNIEAI_RUNTIME_BASE_URL/v1/feedback" \
  -H "Authorization: Bearer $UNIEAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"rating":"good","request_id":"req_from_chat","model":"your-custom-model-slug","comment":"Correct answer","memorize":true,"memory_content":"This product has a two-year warranty."}'

Request fields

Memory is created and synchronized only when all these conditions hold:

1. rating is good.
2. memorize is exactly true.
3. memory_content is non-empty.
4. The Runtime resolves a Customize Model from custom_model_id, model, or request_id.

Successful response:

{
  "success": true,
  "rating": "good",
  "memorized": true,
  "memoryId": "memory_xxx"
}

An accepted rating still returns 200 when no memory is created, but memorized is false and memoryId is null. The Runtime request ID is returned in the x-request-id response header.

Important current feedback limitations

• A bad rating is validated and acknowledged, but this public endpoint currently creates neither a repair item nor model memory.
• When memory conditions are not met, this handler does not separately persist comment as model data.
• Playground thumbs use a separate session-authenticated flow. A thumbs-down can enter the repair queue, while positive content can become pending or explicitly saved memory. Do not assume it is identical to /v1/feedback.
• Memory index synchronization is best-effort. The memory row can exist even if the separate index sync fails; inspect memorized and memoryId and monitor server logs.

Common errors

Security and data handling

• Search results can contain original document text and temporary image URLs. Return them only to authorized callers.
• memory_content can be retrieved in later conversations. Remove passwords, API keys, personal data, and unapproved content before storing it.
• Use separate keys and quotas for each environment, and regenerate a key immediately if exposure is suspected.

Related documentation