Modes

Extraction modes

Instructor supports several ways to extract data from the response.

Output Modes

Instructor supports multiple output modes to allow working with various models depending on their capabilities.

• OutputMode::Json - generate structured output via LLM’s native JSON generation
• OutputMode::JsonSchema - use LLM’s strict JSON Schema mode to enforce JSON Schema
• OutputMode::Tools - use tool calling API to get LLM follow provided schema
• OutputMode::MdJson - use prompting to generate structured output; fallback for the models that do not support JSON generation or tool calling

Additionally, you can use Text and Unrestricted modes to get LLM to generate text output without any structured data extraction.

Those modes are not useful for StructuredOutput class (as it is focused on structured output generation) but can be used with Inference class.

• OutputMode::Text - generate text output
• OutputMode::Unrestricted - generate unrestricted output based on inputs provided by the user (with no enforcement of specific output format)

Example of Using Modes

Mode can be set via parameter of StructuredOutput::create() method.

The default mode is OutputMode::Tools, which leverages OpenAI-style tool calls.

<?php
use Cognesy\Instructor\StructuredOutput;

$structuredOutput = new StructuredOutput();

$response = $structuredOutput->with(
    messages: "...",
    responseModel: ...,
    ...,
    mode: OutputMode::Json
)->get();

Mode, like other parameters can also be set via fluent API methods.

<?php
$response = $structuredOutput
    ->withMessages("...")
    ->withResponseModel(...)
    //...
    ->withOutputMode(OutputMode::Json)
    ->get();

Modes

OutputMode::Tools

This mode is the default one. It uses OpenAI tools to extract data from the response.

It is the most reliable mode, but not all models and API providers support it - check their documentation for more information.

• https://platform.openai.com/docs/guides/function-calling
• https://docs.anthropic.com/en/docs/build-with-claude/tool-use
• https://docs.mistral.ai/capabilities/function_calling/

OutputMode::Json

In this mode Instructor provides response format as JSONSchema and asks LLM to respond with JSON object following provided schema.

It is supported by many open source models and API providers - check their documentation.

See more about JSON mode in:

• https://platform.openai.com/docs/guides/text-generation/json-mode
• https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/increase-consistency
• https://docs.mistral.ai/capabilities/json_mode/

OutputMode::JsonSchema

In contrast to OutputMode::Json which may not always manage to meet the schema requirements, OutputMode::JsonSchema is strict and guarantees the response to be a valid JSON object that matches the provided schema.

It is currently supported only by new OpenAI models (check their docs for details).

NOTE: OpenAI JsonSchema mode does not support optional properties. If you need to have optional properties in your schema, use OutputMode::Tools or OutputMode::Json.

See more about JSONSchema mode in:

• https://platform.openai.com/docs/guides/structured-outputs

OutputMode::MdJson

In this mode Instructor asks LLM to answer with JSON object following provided schema and return answer as Markdown codeblock.

It may improve the results for LLMs that have not been finetuned to respond with JSON as they are likely to be already trained on large amounts of programming docs and have seen a lot of properly formatted JSON objects within MD codeblocks.