Models and ranking
Use a language model for search planning and relevance assessment. Optional preliminary ranking helps choose the candidates for final model scoring.
Connect a model
In Community, or online with your own key, open Settings → Model APIs → Add connection. Choose the connection type, provider, and model, then verify the key, base URL, and protocol. Presets are starting values; the provider must offer that model.
Test connection in the model dialog checks the current draft without saving it. Click Save to create or update the connection, then select it for search or writing. Data-source dialog tests behave differently: they save before testing; see data sources.
| Protocol | Operation |
|---|---|
| OpenAI Chat Completions | /chat/completions |
| OpenAI Responses | /responses |
| Anthropic Messages | /v1/messages |
Use the base URL expected by the protocol, without duplicating the operation path. Chat Completions compatibility does not imply Responses compatibility.
PaperSeek Service provides a model dropdown; PaperSeek manages the endpoints and keys.
Preliminary ranking
Local fusion is the default.
| Mode | Behavior | Extra service |
|---|---|---|
| Local fusion | Combines local lexical/hashed semantic signals and retrieval ranks | None |
| Embedding + Rerank | External embedding and reranking before final selection | Embedding and reranking APIs |
| Decision model | Scores locally fused candidates with the System One protocol | Compatible decision API |
Local fusion refers only to preliminary ranking. The complete search still uses a language model and network requests.
The decision route assesses up to twice the final LLM candidate budget before choosing the final subset. Available candidates and budget limits still apply.
Settings by edition
Community / your own key: configure model routes and credentials. The writing model can differ from the search model.
PaperSeek Service: search uses Auto; review writing uses Fast or Deep Research. Preliminary-ranking choices remain visible, while endpoints and secrets stay on the server.
Local models
Community can connect to a compatible model reachable from the machine running PaperSeek, including loopback.
Online, connections saved with your own key cannot point to loopback, private, link-local, or other unsafe destinations, which protects the server against SSRF. 127.0.0.1 would address the online server, not your laptop. Use Community for local endpoints.
Fallbacks and failures
An optional ranking failure can preserve local ordering with a warning; it does not mean external ranking succeeded. Check the recorded run configuration and execution status.
Authentication, credit, rate-limit, timeout, and invalid-response errors require different remedies. See troubleshooting.