Custom LLM endpoint (BYOK)

Integrava stays hosted in the cloud (embed chat, tickets, workflows, knowledge). You do not install Integrava on your laptop or server.

Instead, you configure an OpenAI-compatible API endpoint in /settings → AI Configuration. When a customer sends a message, Integrava's servers call your URL — the same pattern as pasting an OpenAI API key, but the model runs on infrastructure you choose.

Customer → Integrava cloud (tickets, KB, workflows)
                ↓ HTTPS
         Your LLM endpoint (/v1/chat/completions)
                ↓
         Ollama / vLLM / DeepSeek / etc.

Ollama on your PC listens at http://localhost:11434. That address is only reachable on your machine. Integrava cloud cannot dial your localhost.

To use a self-hosted model with Integrava cloud you need a public HTTPS URL that forwards to your model server, for example:

• Reverse proxy (nginx, Caddy) with TLS on a VPS where Ollama/vLLM runs
• Cloudflare Tunnel, Tailscale Funnel, or similar tunnel to a home lab
• A vendor API (DeepSeek, OpenRouter, Together, etc.) — no tunnel required

In production, Integrava only accepts https:// base URLs and blocks localhost / private IP addresses.

1. Open /settings → AI Configuration.
2. Set Provider to OpenAI (this enables OpenAI-compatible mode even when you are not using OpenAI the company).
3. Set Model to the id your server expects (examples below).
4. Under OpenAI-compatible endpoint, paste your API base URL (Integrava normalizes it to end with /v1).
5. Add an API key if your endpoint requires one. Ollama often does not — leave the key blank or use any short placeholder.
6. Set Embedding model for knowledge search. Your server must support /v1/embeddings with that model name.
7. Save, then click Test connection (or test before saving using the values in the form).
8. Test in Playground or embed chat.

DeepSeek (cloud API)

• Base URL: https://api.deepseek.com/v1
• Model: deepseek-chat
• API key: your DeepSeek key
• Embedding: use DeepSeek embedding model if supported, or keep OpenAI for embeddings only (not mixed today — use one endpoint)

Ollama (self-hosted via HTTPS tunnel)

1. On your server: ollama pull llama3.2 and ollama pull nomic-embed-text
2. Expose https://your-domain.example/v1 to Ollama's OpenAI-compatible port (11434)
3. Model: llama3.2
4. Embedding model: nomic-embed-text
5. API key: optional

OpenRouter

• Base URL: https://openrouter.ai/api/v1
• Model: provider-specific id from OpenRouter (e.g. meta-llama/llama-3.1-8b-instruct)
• API key: OpenRouter key

• Knowledge search needs embeddings from the same compatible endpoint. If your model host has no embeddings API, KB quality will suffer — use a host that supports both chat and embeddings, or a cloud API.
• Vision / attachment OCR requires a vision-capable model on your endpoint. Small local models may not support image inputs.
• Latency depends on your hardware and network path from Integrava to your URL.
• Integrava remains BYOK: you supply URL + key; we do not resell model tokens.

Odysseus is a full self-hosted chat workspace (UI + agents + email). It complements Integrava rather than replacing it. Integrava does not embed Odysseus; if Odysseus (or vLLM/Ollama behind it) exposes a stable OpenAI-style /v1 API over HTTPS, you can point Integrava at that URL using the steps above.