Build RAG on EU GPUs

Upload documents (PDF, TXT, CSV, MD or DOCX) or crawl a URL or sitemap. Your content is split into chunks of about 1000 characters with 100 characters of overlap, on sentence boundaries. Each chunk gets an embedding (384-dimensional). Small files (< 50MB) are fully pre-embedded for fast retrieval; large files use a hybrid mode with FULLTEXT plus on-demand embedding.

There are three ways to attach a knowledge base to your chat requests:

• Link to an instance (recommended): POST /api/kb/{kbId}/link/{instanceId}. Every chat request to that instance then gets context automatically, including via an external hyai- key.
• Link to an agent: set knowledge_base_id on the agent; the public chat uses it automatically.
• Explicit per request: pass knowledge_base_id in the body of your chat completion.

curl https://hostyourai.com/api/v1/chat/completions \
  -H "Authorization: Bearer hyai-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama-3.3-70b",
    "messages": [{"role":"user","content":"What is our return policy?"}],
    "knowledge_base_id": 42
  }'

The Router retrieves the most relevant chunks with cosine similarity (minimum threshold 0.3), injects the top 5 into the prompt, and generates an answer grounded in your documents.

Non-streaming answers include an extra sources array with document name, score, chunk index and an excerpt, so you can trace every answer back to its source. On streaming, context is still injected, but sources cannot be returned in the SSE chunks.

Retrieval failures are graceful: if retrieval fails, the chat continues without context instead of erroring. And because everything runs on EU GPUs, your entire RAG pipeline, from documents to answer, stays inside Europe.

These values determine how your documents are processed and retrieved. They are fixed, so quality stays predictable across all your knowledge bases.