Setup Guide for another AI Provider

This page describes how to configure a generic AI Provider for use with the AI Review Assistant.

If you are using ChatGPT, Claude.ai or Ollama, you can directly consult our dedicated guides.

If any of the headers, body content or query parameters require a secret, use the $API_KEY placeholder, and fill the actual value in the API Key section.

As Repository Admin:

go the Code Review Assistant repository settings page.
open the AI Review Assistant tab on top.

Request section

Section	Instructions
API URL	Enter the REST API address for your AI Provider. This has to be the endpoint for a chat (message based) completion
API Key	Enter your API key
Headers	Add all the key-value Http Headers for your AI Provider. For example, a Basic Authorization or a version controlling header.
Query Parameters	Add all key-value pairs that you would find after the `?` in the url (`?key1=value1&key2=value2`)
Request Body	Add all members for the JSON content. For example, the model, a stop sequence, a maximum number of tokens…

Response Section

Section

Instructions

Success Query

Enter the JSON query to retrieve the content of the chat message completion. For example, if the JSON response from the API is

JSON

{
  "content": [
    {
      "text": "Hi! My name is Claude.",
      "type": "text"
    }
  ]
}

Then the query will be content[0].text

Error Query

Enter the query to retrieve the error message when an error occurs. This is not strictly necessary, but it will help users when an error pops up.

Test and Save

Click on Test and verify that you receive a confirmation message that everything is configured properly.
Click on Save.

If you have trouble in configuring our AI Review Assistant, we are happy to help! 🎉

More on JSON Query Language (JQL)

AI Review Assistant uses a custom JSON Query Language (JQL) to extract text information from JSON formatted responses.

A JQL query is a text snippet, describing a path from the root of a JSON document to a single text field.

Format

A JQL query is made of selections joined with a period (.). For example a.b.c is a path made of three selections (a, b, and c). Selections are evaluated from left to right, and each describes how to extract a nested value from a JSON document. For a JQL query to be valid, the final selection must extract a String value.

Each selection can be one of two kinds, a field selection or an array selection.

Field Selection

A field selection extracts a field of a JSON object. For example message requires the current JSON object to have a field message (e.g. {"message":<value>}) and extracts the associated value.

Array Selection

An array selection combines field selection with indexing. For example content[0] requires the current JSON object to have a field content associated with an array, (e.g. {"content":[<values>...]} ) and tries to extract the element of the array at index 0.

Example

Assume the AI provider returns the following JSON document (this example comes from Claude.ai):

JSON

{
  "content": [
    {
      "text": "Hi! My name is Claude.",
      "type": "text"
    }
  ],
  "id": "msg_013Zva2CMHLNnXjNJJKqJ2EF",
  "model": "claude-3-5-sonnet-20241022",
  "role": "assistant",
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "type": "message",
  "usage": {
    "input_tokens": 2095,
    "output_tokens": 503
  }
}

To extract the useful response message, i.e. "Hi! My name is Claude." - you would use the JQL query content[0].text. First from the root of the document it selects the value of the "content" field (an array) and selects the object at index 0; next from that object it selects the "text" field, resulting in the string we needed.

Similarly, to extract the model used - you would use the JQL query model. From the root of the JSON document it selects the value of the "model" field and extracts the text, i.e. "claude-3-5-sonnet-20241022".