Developers build hallucination-free video chapter generator using Whisper and LLMs

A new open-source pipeline automatically generates video chapters by combining WhisperX transcription with a large language model, avoiding a common AI pitfall. The core problem it solves is that LLMs frequently fabricate timestamps when asked to produce chapters directly from a pasted transcript. The workaround assigns each transcribed segment a numeric index, lets the LLM select chapter boundaries by index only, and then maps those indices back to real timestamps in code. This means the model never handles raw timecodes, eliminating the source of hallucinated times. The pipeline outputs both a WebVTT chapters file and YouTube-compatible timestamp stamps.