Crate response_validator 

Detect and truncate hallucinated conversation turns in LLM responses.

During extended conversations (30+ messages), LLMs can start predicting both sides of dialogue — generating fake [User]: turns within their own responses and then answering them. This crate detects those markers and truncates the response at the first hallucinated turn boundary.

Also detects action hallucinations — when a model claims to have performed actions (file writes, command execution) without corresponding tool calls.

Usage

use response_validator::{validate_text, detect_action_claims};

// Detect hallucinated turn markers
let result = validate_text("Here is my analysis.\n[User]: fake question");
assert!(result.was_truncated);
assert_eq!(result.text, "Here is my analysis.");

// Detect action claims
let claims = detect_action_claims("I've updated MEMORY.md with the notes.");
assert_eq!(claims.len(), 1);

Structs

ActionClaim

A detected action claim in response text.

ActionClaimValidation

Result of validating action claims against actual tool usage.

ValidationResult

Result of validating a response for hallucinated turn markers.

Enums

ActionCategory

Category of action an LLM claims to have performed.

ContentBlock

A content block from an LLM response.

Functions

detect_action_claims

Detect action claims in a text string.

validate_action_claims

Validate action claims in a response against actual tool usage.

validate_content_blocks

Validate content blocks from an LLM response.

validate_text

Validate a text response for hallucinated turn markers.