Module detect 

Hallucination-risk detection from token log probabilities.

The input is the logprobs block that OpenAI-compatible Chat Completions APIs return when you ask for "logprobs": true. Each token’s probability is exp(logprob). Words containing a token whose probability is below the threshold are flagged, and neighbouring flagged words are merged into one span. Each span reports the weakest token and the alternatives the model was weighing at that point.

Low probability is a signal, not proof: a model can be confidently wrong, and it can be unsure about phrasing while the facts are right.

Structs

Alternative

One of the candidates the model considered at a position.

Candidate

An alternative with its probability.

LogprobToken

One generated token with its log probability and the alternatives returned with it.

Report

Result of detect.

Span

A run of low-confidence words.

Constants

DEFAULT_THRESHOLD

Default probability below which a token counts as low confidence.

FLAG_LABEL

Label used for spans produced by detect.

Functions

detect

Flag low-confidence spans.

parse_logprobs

Extract the token list from any of these shapes: a full Chat Completions response (choices[0].logprobs.content), a logprobs object ({"content": [...]}), or a bare token array.

to_token_analysis

Convert tokens plus a report into the visualizer’s input format, so the terminal, HTML and Markdown renderers can display the result.

top_k_entropy_bits

Entropy in bits of alternatives renormalised to sum to 1. None when there are none (or their probabilities sum to 0).