Skip to content
mistral.rs

OpenAI-compatible file inputs

mistral.rs supports OpenAI-compatible user file inputs on the server API and local SDKs:

Use this when a request needs to analyze or transform user-provided files, or when a Responses shell/Skill workflow should see request files in the session working directory.

For field-level compatibility notes, see OpenAI compatibility. For wire schema and lifetime rules, see HTTP API file semantics.

File handling

Section titled “File handling”

file_data is decoded before it reaches the model. Base64 is never placed in prompt context.

Uploaded, inline, URL-fetched, and SDK-provided files are visible through the server file endpoints:

Text-like UTF-8 files get:

  • metadata in prompt context,
  • a decoded preview of up to 4096 chars per file,
  • a total preview budget of 32768 chars per request,
  • access to additional text during agentic tool runs when the preview is not enough.

Binary or non-UTF-8 files get metadata only in prompt context. They are still stored, downloadable from GET /v1/files/{id}/content, and mounted into shell/code workdirs when those tools are active.

Extraction support is intentionally simple:

| File kind | Behavior | |---|---| | Text-like UTF-8 files | Text, CSV, JSON, XML, Markdown, YAML, TOML, HTML, source files, and other valid UTF-8 payloads are readable as text. | | Binary or structured documents | PDFs, images, archives, spreadsheets, and other binary formats are stored and mounted, but mistral.rs does not extract OCR, PDF text, or spreadsheet summaries yet. |

Responses file_url fetches only http and https URLs, with a timeout, redirect cap, decoded-size cap, and basic local/private host rejection. Chat Completions does not support file URLs; upload the file first or use inline file_data.

Start a server:

Terminal window
mistralrs serve --agent -m <model>

--agent is useful when you want file pagination, Python code execution, shell execution, or OpenAI-compatible Skills. Plain requests still receive text previews even without tool execution.

Related setup guides:

Examples

Section titled “Examples”
from openai import OpenAI


client = OpenAI(base_url="http://localhost:1234/v1", api_key="not-used")


with open("data.csv", "rb") as file:
    uploaded = client.files.create(file=file, purpose="user_data")


response = client.responses.create(
    model="default",
    input=[
        {
            "role": "user",
            "content": [
                {"type": "input_file", "file_id": uploaded.id},
                {"type": "input_text", "text": "Summarize this CSV."},
            ],
        }
    ],
)


print(response.output_text)

Full examples:

Other server shapes

Section titled “Other server shapes”

Use inline Responses input_file when you do not want a separate upload:

import base64
from openai import OpenAI


client = OpenAI(base_url="http://localhost:1234/v1", api_key="not-used")


with open("notes.md", "rb") as file:
    data = base64.b64encode(file.read()).decode("utf-8")


response = client.responses.create(
    model="default",
    input=[
        {
            "role": "user",
            "content": [
                {
                    "type": "input_file",
                    "filename": "notes.md",
                    "file_data": f"data:text/markdown;base64,{data}",
                },
                {"type": "input_text", "text": "Extract the action items."},
            ],
        }
    ],
)


print(response.output_text)

Chat Completions uses type: "file" content parts:

import base64
from openai import OpenAI


client = OpenAI(base_url="http://localhost:1234/v1", api_key="not-used")


with open("report.json", "rb") as file:
    data = base64.b64encode(file.read()).decode("utf-8")


completion = client.chat.completions.create(
    model="default",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "file",
                    "file": {
                        "filename": "report.json",
                        "file_data": f"data:application/json;base64,{data}",
                    },
                },
                {"type": "text", "text": "What are the main anomalies?"},
            ],
        }
    ],
)


print(completion.choices[0].message.content)