# Choosing the language model CuratorBot sends one request per interview to a language model, which drafts the summary, chapters, subject headings, journeys and donor checks. The validators then check every quote against the transcript, whichever model wrote it. Each institution chooses where its testimony goes. An administrator sets this once with environment variables (Space secrets and variables on Hugging Face, `.env` locally). Nothing is chosen per user. | `CURATORBOT_LLM_PROVIDER` | Key | Default model | Use it when | | --- | --- | --- | --- | | `openrouter` (default) | `OPENROUTER_API_KEY` | `anthropic/claude-sonnet-5.5`, falling back to `openai/gpt-4.1-mini` | One account reaches many models; good for demos and comparisons | | `anthropic` | `ANTHROPIC_API_KEY` | `claude-sonnet-5-5` | The institution has an Anthropic contract | | `openai` | `OPENAI_API_KEY` | `gpt-4.1-mini` | The institution has an OpenAI contract | | `compatible` | `CURATORBOT_LLM_API_KEY` (if needed) | set `CURATORBOT_LLM_MODEL` | Testimony must not leave the institution: its own gateway, or open models served with vLLM or Ollama | Other settings: - `CURATORBOT_LLM_MODEL`: use a different model from the provider's catalogue. - `CURATORBOT_LLM_FALLBACK_MODEL`: tried once if the main model fails for a reason another model could fix (an outage, a rate limit, an unreadable answer). It is not tried for account problems such as a wrong key or no credit. Set it to an empty value to switch the fallback off. Briefs record which model actually answered. - `CURATORBOT_LLM_BASE_URL`: the address of a `compatible` endpoint, for example `http://localhost:11434/v1` for Ollama. - `CURATORBOT_LLM_JSON_MODE=0`: for local servers that reject the JSON-mode option. - `CURATORBOT_LLM_MAX_TOKENS` (default 16000): the longest answer the model may write. A full interview's brief needs about 6,000; if an answer is cut off, the run falls back to the second model, which the brief records. `/api/health` shows the active provider, model and fallback. ## Data protection - **OpenRouter** forwards each request to the company running the chosen model (Anthropic, OpenAI, or a cloud host of either). In the OpenRouter account's privacy settings, allow only providers that neither train on nor retain prompts before sending pre-accession material. - **Direct contracts** (Anthropic, OpenAI) fall under the institution's own terms with that company. - **`compatible` with a local server** keeps testimony on the institution's own machines. Quality depends on the open model chosen; run the back-test (`docs/real-data-runbook.md`) before relying on it. When a model call fails, curators see a plain sentence ("Live analysis is unavailable right now: …"); the technical detail goes to the server log only.