User Guide
AI Assistant configuration
Configure LLM providers, model selection, and search pipeline settings.
LLM providers#
| Provider | Best for | Requirements |
|---|---|---|
| Azure OpenAI | Enterprise (data residency, compliance) | Azure subscription, endpoint, API key |
| OpenAI | General purpose, fastest to set up | API key |
| Anthropic Claude | Long documents, nuanced answers | API key |
| Ollama | Air-gapped, self-hosted | Ollama server URL, model name |
You can configure different providers for chat (LLM) and embeddings.
Model flexibility#
All model name fields are free-text inputs, not dropdowns. When your LLM provider releases a new model, just type the model name in Settings. No software update, code change, or redeployment needed.
- Azure OpenAI — type your Azure deployment name (e.g. gpt-4o, gpt-5-mini).
- OpenAI — type the model ID (e.g. gpt-4o, o3-mini).
- Anthropic — type the model ID (e.g. claude-sonnet-4-20250514).
- Ollama — type any model you've pulled (e.g. llama3.1, mistral).
Future-proof
The backend passes the model string directly to the provider API with no validation against a model list. Any model your provider accepts will work — today and in the future.
Search pipeline settings#
The Search Pipeline section at the bottom of the AI Assistant page:
- Hybrid Search Balance — 0.0 (pure keyword) to 1.0 (pure semantic). Default 0.70.
- Follow-Up Questions — toggle on/off. Generates 3 related questions after each response.
- Query Reformulation — enable and choose HyDE, Multi-Query, or Step-Back.
- Document Expansion — pull full document context when results cluster.
- Contextual Chunking — prepend document context to chunks (Rule-Based or LLM).
Image search settings#
Enable image search to index and search images using visual embeddings. Choose between OpenCLIP (local, free) or Azure AI Vision (cloud) as the embedding provider. The minimum score threshold controls how strict result filtering is.
Tip
Azure OpenAI is recommended for enterprise deployments. Model fields accept any model your provider supports — no version lock-in.