Add an AI draft you can verify

Keep a useful app when AI is unavailable.

The inbox already works without a model. Prepare rules draft takes an excerpt and adds a fixed follow-up question. It is labelled as rules, because no AI was called. The separate Request AI draft button becomes available only when the server has a key.

To try actual AI, put your own OPENAI_API_KEY in .env and restart Node. Requests use your provider account and may incur charges. Review current model access and pricing first. Only the message text is sent; names or emails typed inside that message would travel with it.

Give the output a shape.

The reference uses the Responses API and a strict schema for a summary, missing-information questions, and a verbatim evidence excerpt. It uses a documented GPT-4.1 mini snapshot that can be changed through OPENAI_MODEL. Read generateSummary and parseSummary in model.mjs.

OPENAI_API_KEY=your_private_key_here
OPENAI_MODEL=gpt-4.1-mini-2025-04-14

These are environment settings, not browser code. Do not paste an actual key into a lesson, screenshot, repository, or shared chat. The API response must be complete and contain a valid draft. A refusal, incomplete response, malformed JSON, or evidence excerpt absent from the original produces an error.

Shape is not truth.

Test an enquiry with no budget or deadline. A useful summary preserves those unknowns. The evidence check verifies an exact excerpt, but it cannot prove every claim in the summary. The original message remains visible for a person to compare.

The request has a 20-second timeout and an output cap. AI attempts are limited to six per minute within this one server process. None of those controls is a guaranteed dollar budget. The source kit’s AI requests are fixture-tested; this release was not tested against a live provider because no key was connected.

Read the official Structured Outputs guide, model documentation, and current pricing before enabling it.