Running FLOW on your own model

Connect FLOW to a model of your own: which APIs it connects to, what your model must offer and how to set it up.

1. FLOW's own models and yours

By default FLOW reads your documents with its own models on its own servers. This is what makes FLOW's privacy and security commitments possible.

Your own model is for organisations with increased security requirements.

2. What changes

With your own model, every read of every document your organisation runs goes to your endpoint instead, and to nothing else. Reading the PDF itself (text and OCR) stays with FLOW.

The owner of the organisation sets the endpoint in Settings. It applies to all runs started after it is saved. Removing it switches back to FLOW's own models.

3. Which APIs FLOW connects to

In Settings, under Model endpoint, choose the API your model is served through: an OpenAI-compatible server (vLLM, llama.cpp server, Ollama, LM Studio and other servers of the OpenAI Chat Completions API), OpenAI, Azure OpenAI, Anthropic (Claude, through its Messages API), or Custom for anything else.

Choosing an API fills in the endpoint URL and the request FLOW sends. Every field stays editable, and FLOW reads with exactly what you save.

4. What your model must offer

Structured output: the model answers in the JSON schema FLOW sends with each request. FLOW asks for exact table rows and figures, and an answer outside the schema is not used.

Token usage in its answers, so FLOW can size its prompts to your context window. Without it FLOW estimates and stays further from the limit.

A context window large enough for a few statement pages and the accounts asked for. We recommend at least 32,000 tokens; a smaller window means fewer pages per read.

An https address on the public internet that answers without redirects.

5. Setting it up

Enter the full endpoint URL (for example https://models.example.com/v1/chat/completions), the model name as your endpoint expects it, the context window in tokens and, if your endpoint should not take more than a number of requests at once, that number.

Enter your API key. The chosen API places it where that API expects it: as a Bearer token, as api-key for Azure OpenAI, as x-api-key for Anthropic. Anything else your endpoint or gateway needs, such as an organisation or routing header, goes under secret headers.

Save and test. FLOW sends one short question with a schema and shows whether your endpoint was reached, whether it answered in the schema, whether it reports token usage and how long it took. Runs are only sent to an endpoint you saved.

6. Advanced: the request

Under Advanced you see and change the request itself: its headers, its body and where FLOW finds the answer. In the body, {{model}}, {{system}}, {{user}}, {{schema}}, {{schemaName}} and {{maxTokens}} stand for FLOW's values; {{apiKey}} can only be used in the URL and the headers. A placeholder FLOW has no value for leaves out the part it sits in, so a request without a schema simply has no schema field.

The answer, the token counts and the stop reason are read with JMESPath expressions, for a whole answer and for a streamed one.

You can add any parameter your API accepts. For reading figures, a temperature of 0 gives the most consistent results: add "temperature": 0 to the body if your API accepts it.

7. Security

FLOW stores your API key and secret header values encrypted and never shows them again; Settings only show their last characters. Only the owner of the organisation can see or change the endpoint.

Your documents go to the endpoint you set. What happens to them there is governed by your agreement with whoever runs that endpoint.

Didn't find what you need? Contact us →