Connect AnythingLLM
Use your Primex key with an existing AnythingLLM desktop or self-hosted workspace. You need access to provider/workspace settings and AI credit for model requests. Choose a chat model in AnythingLLM Setup.
Configure the chat provider
Open Settings → AI Providers → LLM, then choose Generic OpenAI.
| Setting | Value |
|---|---|
| Base URL | https://gateway.primexos.com/v1 |
| API Key | Your dedicated Primex key, entered privately |
| Selected Model | The exact ID selected in Setup |
| Token / context limit | The selected model's published context limit |
| Max Tokens | Start at 1024, or lower if the published output limit requires it |
| Workspace chat Temperature | 1 |
Save the provider. If models load automatically, select the exact ID; otherwise enter it manually. Check any workspace-specific LLM override so your first chat uses this provider/model. If Setup lacks a context limit, confirm it in Catalog before entering a value; do not assume the provider's generic default matches it.
The settings JSON in Setup is a manual checklist, not an AnythingLLM import file. Keep the key in the API Key field. Hosted/server installations need outbound HTTPS access to the Gateway.
First conversation
Start a workspace chat without uploaded documents, agent mode or tools. Send “Reply with a short greeting”, then ask about the answer. Read the full replies and check the exact model and requests in Usage.
AnythingLLM's chat options can change sampling and output limits. Use Temperature 1 for models that reject other values. Inspect Usage before repeating an uncertain request; client retries can create additional requests.
Add documents separately
The LLM provider and Embedding provider are separate settings. You can retain the configured local embedder for an initial document workflow, or configure a Primex embedding model using its embedding contract. Do not choose a chat ID as the embedding model. Changing embedding models can require rebuilding vectors because dimensions and vector meanings differ.
For retrieval, check ingestion, retrieved source passages and the final chat answer as separate stages. Agent, vision and additional provider settings depend on the selected model and their own request contracts.
Fix a connection or remove it
- 401: check the provider key and any workspace override.
- 404: use the
/v1base and the exact model ID, without a duplicated path. - 400: check Temperature, output/context limits and unsupported options.
- Retrieval failure: inspect the embedding provider and ingestion independently.
- 402 / 429 / timeout: use Errors & limits and inspect Usage.
Replace workspace overrides before removing the provider. Revoke the dedicated key once nothing depends on it; provider removal does not remove workspace data.
See the official Generic OpenAI settings source. Report a setup issue.