Structured Extraction From Filings With LLMs

A 10-K buries the number you want — say, guided operating margin for next quarter — inside a paragraph of hedged, lawyer-reviewed prose, on page 40 of 120, phrased differently every filing season. Regex and keyword search break the moment the wording shifts. A large language model reads the sentence like a person would and pulls the number out regardless of phrasing. That's the win. The risk is that it reads the wrong sentence just as fluently, and returns a confident, well-formatted number that is not the one you asked for.

The pattern: narrow the question before you ask it

The reliable approach is not "summarise this filing" but "find this specific field, quote the sentence it came from, and return it in this exact schema." Give the model a JSON schema with named fields — guided_operating_margin_pct, source_quote, fiscal_period — and require every field to be either a value plus the verbatim sentence it was read from, or explicitly null if the filing doesn't state it. Forcing a quote alongside every field turns an unverifiable claim into a checkable one: a human reviewer, or a second automated pass, can confirm the quote actually says what the field claims.

Worked example: guidance extraction

Task: pull forward revenue guidance from an earnings call transcript. The relevant sentence reads: "We expect fiscal year revenue to be in the range of $4.2 billion to $4.4 billion, representing growth of roughly 9 to 14 percent."

A well-specified schema asks for revenue_guidance_low, revenue_guidance_high, currency, fiscal_period, and source_quote. A correctly extracted result: {low: 4200000000, high: 4400000000, currency: "USD", fiscal_period: "FY2026", source_quote: "We expect fiscal year revenue to be in the range of \$4.2 billion to \$4.4 billion..."}. Because the quote is present, a reviewer can check in five seconds that the numbers match the source text — versus re-reading the whole transcript to find where a bare number came from.

The failure mode this guards against: an earlier paragraph in the same call mentions a prior-year comparable figure of $3.9 billion. Without a required quote, a model under time or context pressure occasionally attaches the wrong number to the field — extracting the comparable instead of the guidance. The quote requirement doesn't stop the model from making that mistake, but it makes the mistake visible on inspection instead of invisible in a downstream number.

Building the pipeline

1. Chunk the document by section, not by fixed token count — splitting a guidance sentence across two chunks loses it entirely.
2. Retrieve the relevant chunk rather than feeding the whole filing; see Retrieval-Augmented Generation. This also keeps cost and latency bounded on long documents.
3. Extract with a strict schema and require source quotes for every field.
4. Validate mechanically: does the quote appear verbatim in the source chunk? Does the number parse as the right type? Does the fiscal period match the filing's own metadata?
5. Sample and hand-check a fixed percentage of extractions every run, the same discipline you'd apply to any new data vendor.