Extraction
Extraction turns a message written by a person into facts a computer can compare. One call to a capable language model reads the message together with your profile and returns every lead in it — and, in the same call, its answers to your criteria.
Hi! We’re getting married next April in Collingwood. Around 100–150 guests, sit-down dinner. Call me on 0412 345 678. Sarah
- contact.name
- Sarah Lee
- contact.phones
- 0412 345 678
- request.summary
- Wedding reception for about 120 guests
- location.address
- Collingwood VIC
- fields.event_type
- wedding
- fields.guest_count
- { value: 100, min: 100, max: 150 }
- fields.event_date
- { min: 2027-04-01, max: 2027-04-30, asWritten: "next April" }
Wedding? met — evidence: “We’re getting married next April”
The example message "We're getting married next April in Collingwood. Around 100–150 guests, sit-down dinner. Call me on 0412 345 678. Sarah" becomes: name Sarah Lee, phone 0412 345 678, summary "Wedding reception for about 120 guests", address Collingwood, event type wedding, guest count 100 to 150, event date the whole of next April with the customer's wording kept — and the answer "met" to the wedding question, with the quote "We're getting married next April".
What is extracted
Always, for every lead:
| Part | Contents |
|---|---|
| Contact | Name (and first / last name), emails, phones, other contacts (website, LinkedIn…), the customer's company |
| Request | A one- or two-sentence summary in your workspace language, and the service asked for |
| Location | The address of the job or event, assembled from all its parts, with city, postcode and country |
| Language | The language the customer wrote in |
Plus the fields of your profile — event date, guests, budget, property type, anything — each with a type that decides how the value is read.
Rules the model follows
- Never invent. If the message doesn't say it, the value is empty. A partial value (just a suburb, just a first name) is kept rather than dropped.
- The customer, not the business. Names, phones and emails from your own signature, footer or website are never taken as the customer's. In forwarded mail, the customer is the original sender inside, not the colleague who forwarded it. In form notifications, the customer is whoever filled the form, not the form service.
- Phones as written. Only spaces are removed; no country code or "+" is added.
- The message is data, not instructions. Text inside a message that tries to change the rules or the answers is ignored.
Typed values
Fields come back typed, so that rules can compare them exactly:
- Numbers keep ranges: "100–150 guests" is stored as 100 to 150, and a criterion can look at either end. "Up to 50" and "over 100" are understood.
- Money is an amount with a currency; without a currency in the message, your workspace currency is assumed.
- Dates are resolved from the day the message arrived, in your time zone: "this Saturday", "tomorrow", "the 14th". A date without a year means its next occurrence. Vague periods — "next month", "late March", "in spring" (by your hemisphere), "2027" — become a range of days from the first to the last possible day. "TBA" or "flexible" stays empty. The customer's own wording is always kept.
- Choices (
enum) are matched to your list of allowed values.
Several leads in one message
Some messages contain several enquiries — a daily digest from a lead platform, a staff email listing three callers. Extraction returns one lead per enquiry, each scored on its own. A single enquiry spread over several form fields or quoted replies is one lead.
Answers to your criteria
In the same call, the model answers each AI criterion of your profile: met, not met or unknown, the chosen answer for criteria with several answers, and for every "met" a short quote copied from the message.
Every quote is then checked by code against the message text. A "met" whose quote isn't found in the message is downgraded to unknown. This keeps answers grounded in what the customer actually wrote and defends against messages crafted to trick a model into a higher score.
Enrichment
After extraction, code adds what the message can't say by itself:
- Distance and drive time from your business address to the lead's location (via a maps service), for criteria like "within 30 km".
- Message facts: the channel, the local time it arrived, the day of the week, whether it came in during your business hours.
- Contact facts: the email domains, and how many earlier leads this person has sent.
Each of these can be used in rules. If an enrichment fails — an address that can't be found — the value is simply unknown.