Back to Home

LLM for parsing GOSTs: from 2 hours to 5 min

The article describes the implementation of LLM for automating the extraction of control parameters from GOST scans at a metallurgical enterprise. Control chart preparation time reduced from 2 hours to 5 minutes. Analysis of architecture, prompts, and debugging.

How LLM replaced technologists when working with GOSTs
Advertisement 728x90

Automating GOST Parameter Extraction with LLMs: A 24x Time Reduction Case Study

At a full-cycle metallurgical plant with 4,500 inventory items and over 200 regulatory documents (mostly scans of Soviet-era GOSTs and OSTs), preparing a single control sheet used to take over 2 hours. Technologists manually extracted parameters from tables, notes, and text across 40+ indicators. An LLM-based solution reduced this time to 3–5 minutes—a 24-fold improvement.

The challenge: generate a table with control parameters, sources (section/table), and values based on inventory characteristics (steel grade, billet diameter, group) and a GOST PDF.

Why a Parser Wasn't Enough

A template-based parser failed due to format diversity: in some GOSTs, parameters are in tables; in others, they're in notes or referenced sections. Semantic understanding is required, not just structural parsing.

Google AdInline article slot

The LLM acts as a 'virtual technologist':

  • Input: GOST PDF + inventory parameters
  • Prompt: list of parameters with search algorithms
  • Output: table of 'Parameter — Value — Source'

The Failure of a Universal Prompt

Testing in Perplexity showed: Claude Sonnet 4.6 achieved 85% accuracy on the first GOST, GPT 5.4 achieved 72% (in Thinking mode). But errors recurred on subsequent documents due to structural differences—nested tables, constants, ranges.

A universal approach doesn't scale. Transition to prompts tailored to specific GOSTs.

Google AdInline article slot

Working Architecture Based on the Pareto Principle

80% of inventory is covered by 18% of GOSTs (20 key documents for the pilot).

For each—a custom prompt with rules:

  • Parameter location (table/section)
  • Interpretation of edge cases ('not less than', ranges)
  • Table processing

Process:

Google AdInline article slot
  • Feed GOST PDF and parameters into the prompt.
  • Receive results table.
  • If errors occur—analyze in Perplexity with the command 'Update rule XX'.

9 iterations over 14 days ensured 100% accuracy on pilot GOSTs.

Current Implementation and Scaling

Rules are exported to an Excel table for technologists to edit. The prompt loads the table + input data, outputs a ready table for the enterprise system.

Adding new GOSTs: copy template rules, debug in 1–2 iterations.

Key takeaways:

  • Modern LLMs (Claude Sonnet) reliably parse PDF scans with tables of any complexity.
  • Apply Pareto: start with the top 20% of documents.
  • Custom prompts per document are more effective than universal ones.
  • Iterative debugging eliminates systemic errors (nested tables).

Takeaways for IT Professionals

This approach is applicable in industries with regulatory documentation:

  • Metallurgy, mechanical engineering
  • Chemical industry, construction, energy
  • Pharmaceuticals (without confidential data)

Success conditions:

  • Manual data transfer from heterogeneous PDFs.
  • No strict confidentiality requirements.
  • Ability for iterative prompt refinement.

— Editorial Team

Advertisement 728x90

Read Next