JEDI

Journalistic Excellence through Domain-specific Intelligence

AI-powered transcript processing that cleans up interview transcripts and converts them into polished article format.

The JEDI Project is a collaboration between Atex and Geneea.

---

Installation

pip install -U jedireporter

Quick Start

# Convert plain text interview to transcript JSON
jedireporter-impex import --format plain-text -i interview.txt -o transcript.jsonl

# Process transcript through the full pipeline
jedireporter -i transcript.jsonl -o article.jsonl --llm-profile default

# Export article to markdown
jedireporter-impex export --format markdown -i article.jsonl -o output.md

Configuration

Environment Variables

Copy the template file and fill in your credentials:

cp .env_template .env
# Edit .env with your API keys and credentials

The jedireporter command automatically loads .env from the current directory. Alternatively, set the environment variables directly in your shell.

LLM Profiles

Select an LLM profile using --llm-profile <name>. Available profiles:

Profile	Provider	Model	Environment Variable
default	OpenAI	gpt-5.1-2025-11-13	JEDI_OPENAI_API_KEY
gpt-4.1	OpenAI	gpt-4.1-2025-04-14	JEDI_OPENAI_API_KEY
gpt-5	OpenAI	gpt-5-2025-08-07	JEDI_OPENAI_API_KEY
gpt-5-mini	OpenAI	gpt-5-mini-2025-08-07	JEDI_OPENAI_API_KEY
gpt-5.1	OpenAI	gpt-5.1-2025-11-13	JEDI_OPENAI_API_KEY
gpt-5.2	OpenAI	gpt-5.2-2025-12-11	JEDI_OPENAI_API_KEY
claude	AWS Bedrock	Claude Sonnet 4.5	AWS credentials (see below)
mistral	Mistral	mistral-large-2512	JEDI_MISTRAL_API_KEY

AWS Bedrock Credentials (for claude profile)

export AWS_ACCESS_KEY_ID=<your-access-key>
export AWS_SECRET_ACCESS_KEY=<your-secret-key>
export AWS_DEFAULT_REGION=<aws-region>
export AWS_BEDROCK_ROLE=<bedrock-role-arn>

Langfuse Observability (Optional)

To enable Langfuse tracing for LLM calls, set both environment variables:

export LANGFUSE_PUBLIC_KEY=<your-public-key>
export LANGFUSE_SECRET_KEY=<your-secret-key>

When not configured, the pipeline runs without tracing and no warnings are shown.

Transcript Processing Workflow

The main jedireporter command processes transcripts through a three-stage pipeline:

1. Fix Transcript - Clean up ASR output, fix speaker boundaries, remove fillers
2. Create Article - Convert interview into structured article with topics, Q&A, and summary
3. Post-Process Article - Add styling, enrich with web search results

Usage

jedireporter -i transcript.jsonl -o article.jsonl --llm-profile <profile>

Options

Option	Description
-i, --input-file	Path to input transcript JSON file
-o, --output-file	Path to output article JSON file
--llm-profile	LLM profile name (default: default)
--include-transcript	Include transcript in the article output
--max-concurrency	Maximum parallel workflow invocations (default: 4)

Input Format

Transcript JSON structure (one JSON object per line in .jsonl format).

Minimal Example (mandatory fields only)

{
  "id": "interview-001",
  "language": "en",
  "segments": [
    {"id": "0", "speaker_id": "spk_0", "text": "Welcome to the show."},
    {"id": "1", "speaker_id": "spk_1", "text": "Thank you for having me."}
  ]
}

Full Example (with optional fields)

{
  "id": "interview-001",
  "language": "en",
  "segments": [
    {
      "id": "0",
      "speaker_id": "spk_0",
      "text": "Welcome to the show.",
      "timecodes": {"start_time": 0.0, "end_time": 2.5}
    },
    {
      "id": "1",
      "speaker_id": "spk_1",
      "text": "Thank you for having me.",
      "timecodes": {"start_time": 2.5, "end_time": 4.0}
    }
  ],
  "speakers": [
    {"speaker_id": "spk_0", "role": "host", "name": "John Smith", "description": "Host of the Morning Show"},
    {"speaker_id": "spk_1", "role": "guest", "name": "Jane Doe", "description": "CEO of TechCorp"}
  ],
  "url": "https://youtube.com/watch?v=abc123"
}

Mandatory Fields

Field	Description
id	Unique transcript identifier
language	Language code (e.g., "en", "cs")
segments	List of dialogue segments
segments[].id	Unique segment identifier
segments[].speaker_id	Speaker identifier
segments[].text	Spoken text content

Optional Fields

Field	Description
segments[].timecodes	Start and end times in seconds
speakers	Speaker metadata list
speakers[].speaker_id	Matches segment speaker_id (required if speakers provided)
speakers[].role	"host", "guest", or "other" (required if speakers provided)
speakers[].name	Human-readable name
speakers[].description	Title, affiliation, etc.
url	Source audio/video URL (enables timecode linking in exports)

Import/Export (impex)

Import: Plain Text to Transcript JSON

Convert plain text interviews to transcript JSON format:

jedireporter-impex import --format plain-text -i interview.txt -o transcript.jsonl

Plain text format:

Speaker Name: What they said in this segment...
Another Speaker: Their response goes here...
Speaker Name: And the conversation continues...

Each line follows the pattern Speaker Name: Text content. Empty lines are ignored.

Import: AWS Transcribe JSON to Transcript JSON

Convert AWS Transcribe output to transcript JSON format:

jedireporter-impex import --format aws-transcribe -i transcribe-output.json -o transcript.jsonl

Note: Speaker diarization must be enabled in AWS Transcribe for speaker labels to be included. Without diarization, all segments will be assigned to spk_0.

Options:

Option	Description
--format	Input format: plain-text or aws-transcribe
--language	Language code for the transcript
-i, --input-file	Input file path (supports glob patterns)
-o, --output-file	Output file path

Export: Article JSON to Markdown

jedireporter-impex export --format markdown -i article.jsonl -o output.md

If output is a directory, each article is saved as a separate file.

Export: Article JSON to Google Docs

jedireporter-impex export --format google-doc -i article.jsonl --credentials service-account.json

Options:

Option	Description
--format	Output format: markdown or google-doc
--credentials	Path to Google service account JSON file
--token	Path to OAuth access token file
--google-folder	Google Drive folder ID for output
-i, --input-file	Input article JSON file
-o, --output-file	Output file path (markdown only)

Evaluations

Transcript Fix Evaluation

Evaluates transcript cleanup quality (WER and speaker/segment metrics) by comparing generated transcripts against source ASR and optional gold data.

mkdir -p eval/transcript_fix/results
jedireporter-transcript-eval \
  --generated fixed_transcripts.jsonl \
  --source asr_transcripts.jsonl \
  --gold gold_transcripts.jsonl \
  --output-dir eval/transcript_fix/results

Notes:

• --generated, --source, and --gold are JSONL files of transcript objects (one JSON per line) aligned by id.
• Use --evaluate-missing-gold to continue when gold entries are missing.
• Results are written as per-sample JSON plus an aggregate summary in the output directory (which must exist).

Options:

Option	Description
--generated	Path to JSONL with generated transcripts (skip workflow step)
--gold	Optional path to JSONL with gold transcripts
--source	JSONL with original ASR transcripts used both as workflow input and baseline
--output-dir	Directory to store evaluation results
--evaluate-missing-gold	Run evaluation without matching gold reference instead of failing

Article Pairwise Comparison

Runs an LLM judge that compares two sets of generated articles (model A vs model B). It writes per-pair JSON verdicts plus Markdown reports and can export to Google Docs; a summary.json aggregate is also produced. The required credentials depend on --llm-profile. Set the matching environment variable from the LLM Profiles section.

The judge scores nine criteria:

1. headlines
2. intro & voice
3. interviewee responses
4. interviewer questions
5. readability
6. terminology consistency
7. coverage of transcript content
8. overall reader preference
9. structure/rhythm.

Detailed instructions can be found in the prompt src/jedireporter/eval/article_pairwise_comp/resources/article_pairwise_judge_user.txt.

Each criterion is scored as fully, partly, not_at_all, or not_evaluated; per-criterion winners are article_a, article_b, tie, or not_evaluated.

mkdir -p eval/article_pairwise/results
jedireporter-article-compare \
  --model-a-articles articles_model_a.jsonl \
  --model-b-articles articles_model_b.jsonl \
  --transcripts transcript.jsonl \
  --output-dir eval/article_pairwise/results \
  --llm-profile default \
  --google-api-token token.json

Notes:

• --transcripts is optional; when provided it should align with article IDs (transcript ID without the article- prefix).
• --google-api-token or --google-service-credentials is required to access Google Docs export.
• For automatic Google Docs export with a user account, follow the Google Docs API Python quickstart to generate an OAuth token, then pass the token file to --google-api-token. Update the quickstart script scopes to:

SCOPES = ["https://www.googleapis.com/auth/documents", "https://www.googleapis.com/auth/drive.file"]

• The output directory must exist before running the command.
• Model labels in outputs are inferred from the article JSONL filenames. If the stem contains one of default, gpt-5-mini, gpt-5, mistral, gpt-4.1, or claude, that label is shown; otherwise the full filename stem is used. The default label is displayed as default (gpt-5.1). Options:

Option	Description
--model-a-articles	Path to JSONL with Articles from model A
--model-b-articles	Path to JSONL with Articles from model B
--transcripts	Optional JSONL with source transcripts aligned with the articles
--output-dir	Directory to store evaluation results
--llm-profile	LLM profile name used for article comparison
--max-concurrency	Maximum number of parallel LLM invocations when a batch is provided
--seed	Seed for random number generator. Used for deciding the LLM model for article A/B.
--google-service-credentials	Path to a JSON file with Google API credentials for service account
--google-api-token	Path to a file with access token to Google API
--google-folder	ID of a folder on Google Drive for storing the generated articles to

Example Output (Markdown, short):

Winner: Article A (gpt-5)
Confidence: 0.90
Justification: Model A was better because it won more criteria.

Criterion	Article A (gpt-5)	Article B (mistral)	Better	Justification
1. Headlines	fully	partly	article_a	Model A was better because it met the headline rules.
2. Intro & voice	fully	fully	tie	Both models scored the same.
3. Interviewer questions	partly	fully	article_b	Model B was better because its questions were clearer.
4. Coverage of transcript content	fully	partly	article_a	Model A was better because it covered more topics.

REST API Service

Run the transcript processing pipeline as a web service.

Starting the Server

jedireporter-server --bind localhost --port 5000

Option	Description
--bind	Network interface to listen on (default: localhost)
--port	Port number (default: 5000)

Configuration

The server uses the same environment variables as the CLI:

• JEDI_LLM_PROFILE - LLM profile name (default: default)
• JEDI_SERVER_PATH - URL path prefix for reverse proxy setups

API Endpoints

Endpoint	Method	Description
/status	GET	Health check
/v1/generate-article	POST	Generate article from transcript

Example: Generate Article

curl -X POST http://localhost:5000/v1/generate-article \
  -H "Content-Type: application/json" \
  -d '{
    "id": "interview-001",
    "language": "en",
    "segments": [
      {"id": "0", "speakerId": "host", "text": "Welcome to the show."},
      {"id": "1", "speakerId": "guest", "text": "Thank you for having me."}
    ],
    "speakers": [
      {"speakerId": "host", "role": "host", "name": "John Smith"},
      {"speakerId": "guest", "role": "guest", "name": "Jane Doe"}
    ]
  }'

The API uses camelCase for JSON field names. See /docs for interactive API documentation (Swagger UI).

Command Reference

Command	Description
jedireporter	Main processing workflow (transcript to article)
jedireporter-server	Start the REST API service
jedireporter-impex import	Convert plain text or AWS Transcribe output to transcript JSON
jedireporter-impex export	Convert article JSON to markdown/Google Docs
jedireporter-transcript-eval	Evaluate transcript fix quality against source/gold JSONL
jedireporter-article-compare	Run pairwise article comparison with an LLM judge

Example End-to-End Workflow

# 1. Set up API key
export JEDI_OPENAI_API_KEY=sk-...

# 2. Prepare plain text interview file (interview.txt)
cat > interview.txt << 'EOF'
Host: Welcome to today's show. We have a special guest joining us.
Guest: Thank you for having me. I'm excited to be here.
Host: Let's dive right in. What inspired your latest project?
Guest: It all started when I noticed a gap in the market...
EOF

# 3. Import to transcript JSON
jedireporter-impex import --format plain-text --language en -i interview.txt -o transcript.jsonl

# 4. Process through the pipeline
jedireporter -i transcript.jsonl -o article.jsonl --llm-profile default

# 5. Export to markdown
jedireporter-impex export --format markdown -i article.jsonl -o article.md

# 6. View result
cat article.md

Funding

This project is supported by NextGenerationEU, the European Union's recovery and resilience initiative, through the Recovery and Resilience Facility.

License

This project is licensed under the Apache-2.0 License - see the LICENSE file for details.