Configure services
Open Zotero Settings → Knowledge MCP. The pane has Document parsing, Models and retrieval and MCP connection tabs. Each tab retains its own save or apply action. Switching tabs retains unsaved inputs.
Document parsing
| Parser | Endpoint | Model |
|---|---|---|
| PaddleOCR | https://paddleocr.aistudio-app.com/api/v2/ocr/jobs |
PaddleOCR-VL-1.6 |
| MinerU | https://mineru.net/api/v4 |
vlm |
Select a parser and enter its corresponding API key/token. Enable parsing and save. Each parser retains its own credentials and options when you switch. Models are fixed by the parser selection. Arbitrary OCR endpoints are not automatically compatible with these protocols.
Use HTTPS for external endpoints; plain HTTP is limited to loopback services. Endpoint URLs must not embed passwords, query strings or fragments. An accepted job retains the endpoint and parser identity with which it began; changing the current parser does not redirect that job.
Queue settings control local execution slots, tracked remote jobs, polling, timeouts and file/result limits. Start with defaults. Displayed times use seconds and sizes use MB (1 MB = 1,048,576 bytes).
Models and retrieval
The shipped provider is Bailian. Default models are
qwen3.7-text-embedding and qwen3.7-text-rerank; the default vector dimension
is 768. Verify that your service account supports the selected models, region,
endpoint and dimensions.
- Enter an embedding endpoint, model and key compatible with the selected provider.
- Configure reranking if you want it. Reranking may send retrieved passages to its provider; it has a separate credential or can share the embedding key.
- Use the connection test for each configured service. Tests use the current form values and do not save them.
- Enable services, then Save and apply. Check the applied status.
- Use Vectorize existing documents to generate missing embeddings for previously processed content when needed.
A blank saved-key input preserves the existing key. Connection success, saved configuration and applied runtime state are separate indicators.
Changing an embedding model, endpoint or dimensions changes the vector space. When prompted, choose whether to retain the previous space's vectors or clear them, or cancel the change. Vectors from incompatible spaces are not mixed.
Repair search index rebuilds the local search index from stored vectors without requesting new embeddings. Vectorize existing documents can call the provider and consume quota. See troubleshooting.
MCP connection
Local MCP is optional and disabled by default. Enable it only if a local AI client needs access. Configure the port, apply the connection settings and create a token with explicit knowledge libraries and permissions. Follow the MCP guide.