Model Adapters

A model adapter translates between LF AI Platform’s internal format and the wire format of a specific endpoint. Models that use the custom_connection connection type reference an adapter, which turns the task’s input into an HTTP request body and the endpoint’s response back into an assistant message.

# model_adapters/ragflow.yaml
key: "adapter-ragflow"
display_name: "RAGFlow Chat Completion"
description: "Adapter for RAGFlow models."
process_input:
  language: "jinja"
  source_code: !include "./ragflow_input.jinja"
process_output:
  language: "jinja"
  source_code: !include "./ragflow_output.jinja"

The two transforms

An adapter is a pair of code snippets, each written in Jinja or Python:

  • process_input receives the messages produced by the task’s solver and renders the request body the endpoint expects.
  • process_output receives the endpoint’s response body and renders it in the format LF AI Platform understands, so that scorers see a normal assistant message.

Because the transforms are just code, an adapter can do more than rename fields: it can map tool calls, extract token usage (see Token Usage Tracking), or thread a server-side conversation identifier through a stateful endpoint (see Integrate Stateful Model Endpoint). Snippets are typically kept in their own files and pulled in with !include.

Built-in adapters cover common formats - for example latticeflow$openai_chat_completion and latticeflow$openai_responses - so a custom adapter is only needed for endpoints that deviate from them.

Working with model adapters

Register an adapter with lf add model-adapter and list the available ones, including the built-ins, with lf list model-adapter. A model then references it by key under config.adapter.key. For a step-by-step walkthrough, see Create a Custom Model Adapter.

Configuration

Properties

key Key required

Reference to an existing entity in AI Platform.

Pattern: ^[a-zA-Z0-9_\-\$]+$
Max Length: 250

display_name string required

The model adapter’s name displayed to the user.

description string

Short description of the model adapter.

Default: None

long_description string

Long description of the model adapter. Supports Markdown formatting.

Default: None

provider enum ModelAdapterProviderId

Provider of the model adapter.

Default: user


The provider of the model adapter.

Allowed Values:

  • latticeflow
  • user

process_input ModelAdapterCodeSnippet required

The transform of the model inputs in AI Platform format into the body of the HTTP request.

process_output ModelAdapterCodeSnippet required

The transform of the model’s HTTP response body into the AI Platform format.

display_name: "RAGFlow Chat Completion"
key: "adapter-ragflow"
description: "Adapter for RAGFlow models."
long_description: >
  Adapter for RAGFlow OpenAI compatible messages. See
  [documentation](https://ragflow.io/docs/http_api_reference#openai-compatible-api).
process_input:
  language: "jinja"
  source_code: !include "./ragflow_input.jinja"
process_output:
  language: "jinja"
  source_code: !include "./ragflow_output.jinja"
{
    "model": "{{ model_info.model_key }}",
    "messages": {{ input.messages | tojson }},
    "stream": false,
    "reference": true
}
{% set body = body | fromjson %}
{
    "choices": [
        {% for choice in body.choices %}
            {% set raw_msg = choice.message if choice.message is defined else {} %}
            {% set msg = raw_msg if raw_msg is mapping else {} %}
            {% set clean_msg = {} %}
            {% for k, v in msg.items() %}
                {% if k != "reference" %}
                {% set _ = clean_msg.update({k: v}) %}
                {% endif %}
            {% endfor %}
            {
                "message": {{ clean_msg | tojson }},
                "references": {{ (msg.reference if msg.reference is defined else []) | tojson }}
            }{% if not loop.last %},{% endif %}
        {% endfor %}
    ]
}