audita/README.md

# Audita

Audita is a framework-first transcript correction application. The public `audita` package provides:

- deterministic transcript normalization
- token-batched module orchestration
- concrete `glossary`, `homophones`, `spoken_word`, and `grammar` modules built on reusable proposal / validator contracts
- structured run reporting and work-dir diagnostics

The previous working implementation has been preserved as `audita_prototype` inside this repository. Its full regression suite lives under `tests/audita_prototype`.

## Development

This project is set up for `uv`.

```sh
uv sync --extra dev
uv run pytest
```

## Usage

Process a transcript with the current framework implementation:

```sh
uv run audita process transcript.json --glossary glossary.yaml --output corrected.json
```

The framework currently runs this default module sequence:

1. `glossary`
2. `homophones`
3. `glossary`
4. `spoken_word`
5. `grammar`

Resolved run instance names are auto-numbered for repeats, so the default report pipeline is:

1. `glossary_1`
2. `homophones`
3. `glossary_2`
4. `spoken_word`
5. `grammar`

The default module sequence is fully implemented today:

- `glossary` proposes glossary-supported acoustic corrections
- `homophones` proposes conservative homophone and mistranscription corrections
- `spoken_word` proposes conservative dysfluency cleanup
- `grammar` proposes punctuation, capitalization, and spacing cleanup only

To run a custom module sequence, pass `--modules`:

```sh
uv run audita process transcript.json --glossary glossary.yaml --modules grammar --output corrected.json
```

To also write a structured JSON report:

```sh
uv run audita process transcript.json --glossary glossary.yaml --output corrected.json --report-json report.json
```

From a checked-out repository, you can also use the root launcher:

```sh
./audita process transcript.json --glossary glossary.yaml --output corrected.json
```

For a system-wide command, install the source tree under `/usr/local/src/audita`, sync dependencies there, and symlink the root launcher into your `PATH`:

```sh
cd /usr/local/src/audita
uv sync --extra dev
ln -s /usr/local/src/audita/audita /usr/local/bin/audita
audita process transcript.json --glossary glossary.yaml --output corrected.json
```

Without `--output`, Audita writes the corrected transcript JSON to stdout and progress logs to stderr.
`--report-json` writes a separate machine-readable run report and never mixes report data into stdout.

Useful configuration can be supplied by CLI flag or environment variable. CLI flags take precedence over environment variables. Default OpenRouter runs require LLM API credentials, because the `glossary`, `homophones`, `spoken_word`, and `grammar` modules make real LLM calls. Self-hosted or other non-default OpenAI-compatible endpoints may not require credentials. `AUDITA_LLM_API_KEY` and `--llm-api-key` are the preferred provider-neutral credential surfaces, while `OPENROUTER_API_KEY` remains supported as a backward-compatible fallback.

| Environment variable | CLI flag | Default | Purpose |
| --- | --- | --- | --- |
| `AUDITA_MODULES` | `--modules` | `glossary,homophones,glossary,spoken_word,grammar` | Comma-separated logical module keys to run; CLI overrides the environment value |
| `AUDITA_LLM_API_KEY` | `--llm-api-key` | unset | Preferred provider-neutral LLM API credential; required for the default OpenRouter endpoint and optional for non-default endpoints; CLI overrides both environment-key variants |
| `AUDITA_VALIDATION_LLM_API_KEY` | `--validation-llm-api-key` | unset | Validation-phase LLM API credential; defaults to the primary LLM API key and is optional for non-default validation endpoints |
| `AUDITA_MODEL` | `--model` | `openrouter/google/gemma-4-31b-it` | LLM model name sent to the configured OpenAI-compatible endpoint |
| `AUDITA_VALIDATION_MODEL` | `--validation-model` | unset | Validation-phase LLM model; defaults to `AUDITA_MODEL` |
| `AUDITA_BASE_URL` | `--base-url` | `https://openrouter.ai/api/v1` | OpenAI-compatible API base URL |
| `AUDITA_VALIDATION_BASE_URL` | `--validation-base-url` | unset | Validation-phase OpenAI-compatible API base URL; defaults to `AUDITA_BASE_URL` |
| `AUDITA_LLM_TIMEOUT_SECONDS` | `--llm-timeout-seconds` | `600` | Per-request timeout in seconds for LLM calls to the configured OpenAI-compatible endpoint |
| `AUDITA_VALIDATION_LLM_TIMEOUT_SECONDS` | `--validation-llm-timeout-seconds` | unset | Validation-phase per-request timeout in seconds; defaults to `AUDITA_LLM_TIMEOUT_SECONDS` |
| `AUDITA_VALIDATION_MAX_PROMPT_TOKENS` | `--validation-max-prompt-tokens` | `2048` | Maximum estimated tokens per validation-phase LLM prompt batch |
| `AUDITA_TARGET_SECTIONS` | `--target-sections` | unset | Exact number of contiguous proposal-stage transcript sections; errors if min/max token bounds cannot be satisfied |
| `AUDITA_MAX_RETRIES` | `--max-retries` | `3` | Maximum Instructor retries for structured responses |
| `AUDITA_VALIDATION_MAX_RETRIES` | `--validation-max-retries` | unset | Validation-phase structured-output retries; defaults to `AUDITA_MAX_RETRIES` |
| `AUDITA_VALIDATION_LLM_CONCURRENCY` | `--validation-llm-concurrency` | unset | Validation-phase LLM concurrency; defaults to `AUDITA_LLM_CONCURRENCY` |
| `AUDITA_MAX_SECTION_TOKENS` | `--max-section-tokens` | `8192` | Maximum estimated tokens per proposal-stage transcript section |
| `AUDITA_MIN_SECTION_TOKENS` | `--min-section-tokens` | `2048` | Minimum estimated tokens per proposal-stage transcript section when balancing for concurrency |
| `AUDITA_GLOSSARY_CONFIDENCE_THRESHOLD` | `--glossary-confidence-threshold` | `0.8` | Minimum confidence required for glossary proposals to survive validation |
| `AUDITA_GRAMMAR_CONFIDENCE_THRESHOLD` | `--grammar-confidence-threshold` | `0.8` | Minimum confidence required for grammar proposals to survive validation |
| `AUDITA_HOMOPHONES_CONFIDENCE_THRESHOLD` | `--homophones-confidence-threshold` | `0.8` | Minimum confidence required for homophone proposals to survive validation |
| `AUDITA_SPOKEN_WORD_CONFIDENCE_THRESHOLD` | `--spoken-word-confidence-threshold` | `0.8` | Minimum confidence required for spoken-word proposals to survive validation |
| `AUDITA_NORMALIZE_MAX_SEGMENT_GAP` | `--normalize-max-segment-gap` | `4.0` | Same-speaker gaps eligible for deterministic merging |
| `AUDITA_NORMALIZE_ELLIPSIS_GAP` | `--normalize-ellipsis-gap` | `3.5` | Same-speaker gaps above this value are joined with ` ... ` |
| `AUDITA_NORMALIZE_MAX_SEGMENT_DURATION` | `--normalize-max-segment-duration` | `60.0` | Maximum merged segment duration |
| `AUDITA_NORMALIZE_MAX_SEGMENT_TOKENS` | `--normalize-max-segment-tokens` | `2048` | Maximum merged segment prompt payload size |
| `AUDITA_WORK_DIR` | `--work-dir` | `/tmp/audita` | Per-run scratch diagnostics directory |
| `AUDITA_WORK_DIR_RETENTION` | `--work-dir-retention` | `auto` | Whether to retain the per-run work directory: `auto`, `always`, or `never` |

Set `AUDITA_MODULES=grammar` to run only the grammar module by default, or override it per command with `--modules`.

Validation-phase LLM settings inherit from the primary `AUDITA_*` LLM settings by default. Set any of the `AUDITA_VALIDATION_*` values only when you want LLM-backed validators to use a different model, endpoint, credential, timeout, retry budget, or concurrency level.

OpenRouter remains the default out of the box:

```sh
export AUDITA_LLM_API_KEY=your-openrouter-key
audita process transcript.json --glossary glossary.yaml --output corrected.json
```

You can point Audita at any OpenAI-compatible endpoint by changing `AUDITA_BASE_URL` and, if needed, `AUDITA_MODEL`. For example, a local vLLM server:

```sh
export AUDITA_BASE_URL=http://localhost:8000/v1
export AUDITA_MODEL=meta-llama/Llama-3.1-8B-Instruct
audita process transcript.json --glossary glossary.yaml --output corrected.json
```

If your self-hosted endpoint requires authentication, you can still set `AUDITA_LLM_API_KEY`; Audita simply no longer requires it for non-default endpoints.

Or the actual OpenAI API:

```sh
export AUDITA_LLM_API_KEY=your-openai-key
export AUDITA_BASE_URL=https://api.openai.com/v1
export AUDITA_MODEL=gpt-4.1-mini
audita process transcript.json --glossary glossary.yaml --output corrected.json
```

`AUDITA_WORK_DIR` stores per-run diagnostics while processing. Under the default `AUDITA_WORK_DIR_RETENTION=auto`, clean successful runs are removed, while failed runs and successful runs with final skipped corrections are preserved. Use `always` to keep every run directory and `never` to remove successful run directories even when skips remain.
Failed runs always preserve the run directory and include an authoritative `report.json` alongside normalization and prompt/response diagnostics.

## Prototype Archive

The archived prototype remains importable as `audita_prototype` and is still covered by its original regression suite. This is intentional: the new `audita` package is a framework-oriented rewrite, not a thin wrapper around the old code.