In short
Turn messy text like visiting cards, invoices or job posts into clean rows in a spreadsheet. This project teaches structured output, validation and retries, which are core skills for real LLM work.
A tool that pulls structured fields (names, dates, amounts) out of unstructured text like invoices or emails — a common, high-value real-world LLM application.
Think of it like this
Filing the visiting cards
After a college fest, you have a heap of visiting cards in every design. You read each one and file it: name, company, phone, email, in the same order every time. Your extractor does this for any messy text, at scale.
Watch it happen
One document through the extractor
job_post.txt
1/6A user uploads a messy job post copied from a website.
The big ideas
What you need to know
Schema first
Decide the exact fields and types before writing prompts. The schema is the contract.
Validate and retry
Validate every result. Retry once with the error message, then flag for human review.
Nulls over guesses
If a salary isn't mentioned, it should be null, not a made-up number.
Measure accuracy
Hand-label 20 documents and compare field by field. Report accuracy per field, not just overall.
See it in code
Walk through the code
1from pydantic import BaseModel, ValidationError2 3class JobPost(BaseModel):4 title: str5 company: str6 city: str | None7 salary_lpa: float | None # lakhs per annum, null if not stated8 skills: list[str]9 10def extract(text: str) -> JobPost | None:11 prompt = f"Return only JSON for this schema: {JobPost.model_json_schema()}\n<text>{text}</text>"12 for attempt in range(2):13 try:14 return JobPost.model_validate_json(ask_llm(prompt))15 except ValidationError as e:16 prompt += f"\nYour last JSON was invalid: {e}. Fix it."17 return None # send to human review1/3The schema with optional fields where data might be missing.
Schema, prompt, validate, retry once, then give up gracefully. ask_llm is your own helper.
Save yourself the pain
Common mistakes
Oops
Only testing on clean, well-formatted examples.
Do this instead
Include messy ones: missing fields, two salaries, Hinglish, extra emojis. That's what real data looks like.
Oops
Retrying forever when validation fails.
Do this instead
Retry once or twice, then flag for review. Endless retries waste money and hide real problems.
Your mission
First milestone: 10 rows
Quick check
A job post doesn't mention salary. What should your extractor return for salary_lpa?
Go deeper
- 1Define a strict output schema and use structured-output mode to extract fields reliably.
- 2Test against messy, real-world input text, not just clean examples.
- 3Add validation and a fallback path for fields the model couldn't confidently extract.
More projects at this stop
AI Writing Assistant
Stuck on this topic? Ask an Insider
Get 1:1 guidance from people who've walked this exact path — on the InsideEdge app.
