Providers
The app supports OpenAI, Vercel AI Gateway, xAI, Ollama Cloud, OpenRouter, and custom OpenAI-compatible endpoints. Fella is BYOK-only. See Models and providers for connection steps and the defaults in the current source.Model selection
The user selects a model for each conversation from the composer. Provider credentials are stored at the app level. Fella does not automatically route an Investigation across different providers or models. When a consistency check is enabled and triggered, it uses the same selected provider and model. This keeps model choice separate from the analytics engine. The model can direct tool use, but the Rust engine decides whether a tool call is valid and performs the local computation.Request and response path
- The harness sends the selected provider the user request, relevant conversation and workspace context, and the schemas for available tools.
- The provider returns text, structured tool calls, or a final answer.
- The Rust harness validates and dispatches tool calls; the engine returns bounded results.
- The answer and evidence are saved locally with the conversation and analysis run.
- When useful result data is available, the model composes a JSON-render view that references those saved results.
Generated presentation boundary
JSON-render is the current structured answer renderer. The model composes from Fella’s registered component catalog, which includes prose, layout, metrics, tables, charts, callouts, and presentation controls. The model may choose different layouts for different questions. Presentation output is declarative data, not executable markup or code. Analytical values must reference supplied result IDs and fields. Validation checks the component tree, props, and result bindings before the Svelte renderer mounts it. A presentation error does not invalidate the canonical answer or its evidence. The component catalog is a safety and consistency boundary. It does not mean every response must contain a chart, nor does it prescribe one fixed table layout.Data sent to providers
The selected provider may receive the question, recent conversation context, workspace context, tool schemas, and requested tool results. Results can include file-derived text, table rows, aggregates, or Python output. Fella does not upload the whole folder as a single file. Read Privacy and security before analyzing sensitive data.Limits
- Provider model catalogs and identifiers can change; the composer’s current list is authoritative when available.
- A model can select the wrong source or misunderstand the question.
- Evidence and deterministic checks can expose some errors but do not prove semantic correctness.
- Fella does not provide a hosted inference service or automatic multi-model routing.
- The inert /mcp command does not add model tools or connector traffic.
The harness
Follow context, tool calls, evidence, and presentation generation.
Privacy and security
Review provider egress, credentials, and local execution.