regex-vs-llm-structured-text

CIEL ADAPTATION: Regex vs. LLM (Parsing Strategy)

This skill provides a decision framework for structured text extraction, prioritizing deterministic speed (Regex) for the common case and reserving LLM reasoning for the long tail.

The Hybrid Pipeline Pattern

1. Regex Parser (95%): Attempt to extract structure using deterministic regular expressions. Optimized for speed and cost.
2. Confidence Scorer: Programmatically flag items that deviate from expected patterns (e.g., missing fields, short text).
3. LLM Validator (5%): Dispatch low-confidence items to a cheap LLM (e.g., Haiku) for intelligent extraction/fixing.

Decision Matrix

• Use Regex IF: The format is consistent (>90% repeating), and you need deterministic speed.
• Use LLM IF: The text is free-form, highly variable, or the regex parser returns low confidence (<0.95).

Efficiency Metrics (Target)

• Cost Savings: >90% reduction compared to an all-LLM pipeline.
• Accuracy: >99% combined accuracy (Regex speed + LLM reasoning).
• Latency: Near-instant processing for the majority of inputs.

Implementation Standards

• Chunking: For large documents, chunk the input to prevent regex timeouts.
• Test-Driven Development: Write tests for known patterns first, then add LLM fallbacks for the edge cases.
• Statelessness: Parsers should be pure functions that do not mutate input state.

Anti-Patterns

• Brute Force LLM: Sending simple repeating forms to an LLM without an initial regex attempt.
• Regex Purity: Attempting to write a 500-line regex to handle an edge case that an LLM can solve in one sentence.
• Blind Faith: Assuming a regex "just works" without a programmatic confidence/validation check.