Skip to main content
NoteWise sends generation requests through LiteLLM. Use a model string like provider/model-name with --model or DEFAULT_MODEL.
The source default is gemini/gemini-3.5-flash.

​
Common routes

Default Gemini path

Set GEMINI_API_KEY and use gemini/gemini-3.5-flash.

OpenAI or Anthropic

Set OPENAI_API_KEY or ANTHROPIC_API_KEY, then use a matching LiteLLM model string.

Hosted routing

Use OpenRouter, Vercel AI Gateway, Azure, Bedrock, Databricks, Cloudflare, Together, Fireworks, or Groq.

Subscription login

Use ChatGPT or GitHub Copilot OAuth instead of static API keys.

Custom endpoint

Save one or more OpenAI-compatible endpoints with notewise setup or notewise inference.

​
Credential map

Never share API keys, OAuth tokens, cookies, raw prompts, or provider payloads. Logs and error reports should stay redacted.

​
Model inference

When a model string has no explicit provider prefix, NoteWise can infer common key requirements from model names such as Gemini, GPT/OpenAI, Claude, Groq, Grok/xAI, Mistral, Cohere/Command, and DeepSeek.

​
Saved custom endpoints

notewise process --model <name>/<model-id> automatically resolves a saved endpoint named <name>, uses that profile’s base URL and API key, and sends LiteLLM openai/<model-id>. This lookup is exact after endpoint-name normalization, so profiles remain isolated. Use explicit --base-url or --api-key only to override a single run; neither flag changes the saved profile. When --base-url changes a saved endpoint, supply --api-key so the saved credential is never sent to a different origin. Use notewise inference list, add, update, and delete to manage the saved profiles. add and update are interactive: any argument left out is prompted for, and an omitted --model opens the same paginated, searchable model picker used by notewise setup. See Commands for the exact prompted flags, and Configuration for the registry format and setup behavior.

​
Per-token pricing for custom endpoints

When NoteWise discovers and verifies a model during notewise setup, notewise inference add, or notewise inference update, it also reads the endpoint’s /v1/models response for per-token prompt/completion pricing, trying several real-world response conventions. Any pricing found is saved per endpoint and model alongside the profile in config.db. At generation time, NoteWise always tries LiteLLM’s own built-in cost catalog first. Only when that catalog has no entry for the model does it fall back to this saved pricing, routed through LiteLLM’s own cost calculation so a custom rate is priced the same way as any other model. This means notewise stats and notewise history can show real costs for custom or gateway models that previously always showed $0. Endpoints that don’t advertise pricing keep showing $0 estimated cost, exactly as before.
Last modified on November 6, 2026