How to Build Custom Question Profiles for Smarter Document Extraction

What Are Question Profiles?

A question profile is a reusable set of structured questions that tells an AI extraction platform exactly what data to pull from your documents. Think of it as a checklist for the AI — instead of returning everything it finds, it answers the specific questions your team actually needs answered.

Each question in a profile includes three elements: the question text itself, the expected answer type (text, number, date, yes/no, or list), and example answers that guide the AI toward the right format and level of detail.

The principle is straightforward: the more precisely you define what you need, the more accurate the results. Vague questions produce vague answers. Specific, well-typed questions with clear examples produce structured data you can act on immediately.

Why Profiles Matter: Garbage In, Garbage Out

Consider two approaches to extracting financial data from a contract:

Weak approach: "What are the financial terms?" This produces a wall of text summarizing anything related to money — all mixed together in a narrative paragraph.

Strong approach: Five separate questions, each with a specific type:

• "What is the total contract value?" (number) — Example: "EUR 2,500,000"
• "What are the payment milestones?" (list) — Example: "30% on signing, 40% on Phase 1, 30% on acceptance"
• "What is the retention percentage?" (number) — Example: "5%"
• "What penalty applies for late delivery?" (text) — Example: "0.1% per day, capped at 10%"
• "What is the payment term?" (text) — Example: "Net 30 days from invoice date"

The second approach produces five clean data points that can be tabulated, compared across contracts, and exported to a spreadsheet.

The Pre-Built Profiles: A Starting Point

extriq ships with five industry-specific question profiles:

• Construction / Building — Project identification, deadlines, budget, materials, warranties, insurance, compliance
• Legal / Contracts — Parties, dates, termination, liability, indemnification, governing law, disputes
• Finance / Invoices — Vendor details, invoice numbers, line items, tax, payment terms, bank details
• Tenders / Procurement — Submission requirements, evaluation criteria, bid bonds, eligibility, specifications
• General Document Analysis — Metadata, key parties, dates, financial figures, obligations, conditions

These are designed to deliver value out of the box and serve as a foundation for customization.

How to Create a Custom Profile

Step 1: Identify the document type. Insurance policies, environmental assessments, supplier qualification forms? The more specific, the more targeted your questions.

Step 2: List the data points your team needs. Interview the people who currently review these documents manually. What do they look for? What do they enter into spreadsheets?

Step 3: Assign answer types. Each question should have one of five types:

• Text — Names, descriptions, clauses. Example: "Who is the contracting authority?"
• Number — Values, quantities, percentages. Example: "What is the bid bond amount?"
• Date — Deadlines, effective dates. Example: "What is the contract start date?"
• Yes/No — Binary conditions. Example: "Does the contract include a force majeure clause?"
• List — Multiple items. Example: "What certifications are required?"

Step 4: Write example answers. This is the most important step. Examples tell the AI what a correct answer looks like. Provide 2-3 examples per question, ideally from real documents.

Step 5: Order and group questions logically. Arrange questions in the order a human reviewer would work through the document. Group related questions together.

Tips for Writing Effective Questions

• Be specific, not broad. Instead of "What are the project requirements?" ask "What technical certifications must the contractor hold?"
• Match the question to the answer type. "When must the bid be submitted?" (date) is better than "What are the submission requirements?" (text)
• Use domain terminology. "What is the LD cap?" will extract liquidated damages caps. "What is the SLA response time?" will find service level metrics.
• Include edge case examples. If a date might appear in multiple formats, include examples of each.
• Keep questions independent. Each question should stand on its own — avoid references to previous answers.

Common Mistakes to Avoid

Asking compound questions. "What is the project value and timeline?" combines two data points. Split it into two questions with different answer types.

Using the wrong answer type. Asking "What is the contract value?" as text produces "approximately two and a half million euros." Using number produces "EUR 2,500,000."

Writing questions that are too open-ended. "Summarize the document" is not an extraction question.

Skipping examples. Questions without examples force the AI to guess what format you expect.

Duplicating pre-built profile questions. If a pre-built profile covers 80% of your needs, duplicate it and add custom questions rather than starting from scratch.

Putting It Into Practice

The best way to develop profiles is iteratively. Start with the closest pre-built profile, run an extraction on sample documents, review results, and refine. Adjust wording when answers are imprecise, add examples when format is inconsistent, and split questions when answers cover too much ground.

Over time, your custom profiles become institutional knowledge — a codified version of what your team looks for. New team members use the same profiles and get consistent results from day one.

Ready to Get Started?

Try extriq free for 30 days and build your first custom question profile. Start with one of the five pre-built profiles and customize it for your specific document type.

Start your free trial