LLM Queries¶

These functions send prompts to language models and return structured responses.

query_model¶

query_model(model, prompt, context)

Queries a declared model with a prompt and optional context.

Parameters:

• model - A model variable (created via declare ... as model)
• prompt - A string containing the question or prompt
• context - An optional record providing additional context for the query

Returns: A record whose structure depends on the model's configuration:

With context (binary mode):

• "value" - Boolean answer
• "p_yes" - Probability of "yes"
• "p_no" - Probability of "no"
• "confidence" - Confidence score

With response_schema configured:

Returns a record matching the schema defined in the model declaration.

Without context (free-form mode):

• "text" - The raw LLM response text
• "confidence" - Always 1.0

Example: binary question answering¶

declare my_model as model {
    model_name = "chat",
    temperature = 0.7
}

let response = query_model(
    my_model,
    "Is climate change caused by human activity?",
    {"topic": "science", "source": "IPCC"}
);

let answer = get(response, "value");
let confidence = get(response, "confidence");

Example: collecting multiple responses¶

declare model as model { model_name = "chat", temperature = 0.8 }
let responses = [];

for 1..20 {
    let result = query_model(model, "Is AI beneficial?", {"topic": "ai"});
    append responses with result;
}

let values = get_all(responses, "value");
let consensus = majority_vote(values);
let agreement = agreement_rate(values);

Example: custom response schema¶

declare classifier as model {
    model_name = "chat",
    temperature = 0.3,
    response_format = "json",
    response_schema = {
        "type": "object",
        "properties": {
            "category": { "type": "string" },
            "confidence": { "type": "number" }
        }
    }
}

let result = query_model(classifier, "Classify this text: 'I love this product'");
// result matches the defined schema

llm_query¶

llm_query(prompt, context)

Queries the default LLM with a prompt and optional context. Unlike query_model, this function does not require a declared model.

Parameters:

• prompt - A string containing the question or prompt
• context - An optional record providing additional context

Returns:

With context:

A record in binary mode with "answer", "confidence", "p_yes", and "p_no" fields.

Without context:

A record with "answer" (the raw text response) and "confidence" (always 1.0).

let response = llm_query("What is the capital of France?");
let answer = get(response, "answer");