at-field
at-field measures how much of a theme's vocabulary a transcript uses. Give it a transcript and a theme, for example --theme "animals", and it returns a score with the matches behind it.
at-field transcript.srt --theme "economy"
A replay of two real runs. The model's wait and the report path are trimmed.
Transcripts can come from any tool: whisper.cpp, a transcription service, YouTube captions. at-field starts from the text.
Reading the score
Every run reports a lexical saturation score: the share of the theme's vocabulary that shows up in the transcript (distinct field words found ÷ field size), with the raw counts and a matches-per-1,000-words density beside it. It works best on a theme that sits beneath the transcript's subject, such as economy in a nature documentary or conflict in a cooking show.
The report says whether the score falls below, between, or above two boundaries (default 10% and 70%). The defaults are arbitrary starting points; set your own with --saturation-low and --saturation-high. What a position means is up to you.
Install
Requires Node.js 22+ and Ollama running locally. Ollama expands the theme into a lexical field; nothing is bundled and no API key is needed.
ollama pull qwen2.5:3b
npm install -g at-field
The default model is about 1.9 GB and downloads once. If it is missing, at-field shows its size and offers to pull it. A run with no terminal (a script or CI job) starts no download and prints the ollama pull command instead.
To try it without installing, run it through npx:
npx at-field transcript.srt --theme "economy"
npx fetches the package on first use, so the first start takes longer. Set up Ollama and pull the model first, because npx fetches only at-field.
To run from source instead:
git clone https://github.com/promethee/at-field.git
cd at-field
npm install
npm run build
npm link
at-field talks to Ollama's default local server (http://localhost:11434). Set the OLLAMA_HOST env var if yours runs elsewhere.
Usage
at-field <transcript-file> --theme "<theme>" [options]
Accepts .srt and .vtt (timestamps are kept in the report) or plain text (no timestamps).
| Flag | Default | Description |
|---|
--theme <value> | (required) | Theme to expand into a lexical field and score against the transcript. |
--lexic <path> | none | Wordlist file to use instead of model expansion, used as written. --theme still names the analysis. |
--language <code> | detected from the transcript | Transcript's language, for example en or fr. The lexical field is written in this language. |
--model <name> | qwen2.5:3b | Ollama model for theme expansion. Offers to ollama pull it when missing. Smaller models such as qwen2.5:0.5b run faster and give weaker fields outside English. Larger ones such as mistral may do better in French. |
--field-size <n> | 25 | How many words to keep when the model expands the theme (5 to 100). Raise it for long transcripts. Cannot be combined with --lexic. |
--stem-length <n> | 5 | Leading letters that identify a word's stem, from 3 to 12, or 0 to turn stem filtering off. Words sharing a stem count as one word ("voyage" and "voyageur"), and words sharing a stem with the theme are dropped ("economic" for "economy"). Lower it to merge more forms, raise it to merge fewer. Cannot be combined with --lexic. |
--saturation-low <pct> | 10 | Low boundary for the saturation score, in percent. |
--saturation-high <pct> | 70 | High boundary for the saturation score, in percent. |
--verbose | off | Also print the full Markdown report and the long limitation notices. |
--quiet | off | Print only the report path. Cannot be combined with --verbose. |
--version | off | Print the version and exit. |
Each run writes a Markdown report next to the input transcript.
Philosophy
Limitations that apply to a run (missing timestamps, a thin field) print as one notes line, and the report file carries the full disclaimers, including that accuracy depends on the tool that made the transcript. Progress and notes go to stderr, so stdout can be piped. Theme expansion runs through a local Ollama server, with no cloud API calls and no per-run cost. Output is Markdown.
Scope
at-field measures one theme in one transcript per run and writes one report. Other jobs belong to other tools: a summarizer for what a text covers, a transcription tool such as Whisper or Vibe for audio, and a topic-modelling tool for discovering themes. It keeps no project state between runs.
Known limitations
- The lexical field is written in the transcript's language (detected, or set with
--language), so --theme can be in any language. Field quality depends on the model, and small models are weak outside English, which is why qwen2.5:3b is the default.
- Matching is literal. A word matches only in the form listed, so "jeu" and "jeux" count separately. Include the forms you want through
--lexic.
- Plain-text input carries no timing, so the timestamped-occurrence log is empty. Use
.srt or .vtt input to keep it.
- A field under 8 words triggers a thin-field notice, since the result then reads more like keyword spotting than a broad thematic measure.
Performance
Theme expansion is the slow step, and its speed depends on the hardware running Ollama. Ollama uses a supported GPU automatically, and a GPU is recommended if you run at-field often. On a CPU-only machine, the default 3B model took about 2 to 5 minutes per run in testing. A smaller model (--model qwen2.5:0.5b) runs faster and gives weaker fields outside English. A wordlist (--lexic) skips expansion and answers in seconds.
License
MIT
Made with AI, designed by human
Built through pairing with Claude Code. A human made every design decision, trade-off and rejected alternative, all recorded in INTENT.md; the AI handled implementation.
