Parsing Messy Language into DSL: Stability Challenges with DeepSeek V4 Flash

Ok_Pop3337 · reddit · 2026-08-25

A developer shares experience building a conversational finance tool that converts natural language to typed financial operations using DeepSeek V4 Flash. While deterministic validation works safely, the LLM struggles with semantic stability: dropping plausible interpretations, merging distinct operations, and hallucinating fields like dates. Attempts using stricter schemas, larger prompts, high reasoning effort, and separate judging steps yielded limited improvements. The author seeks advice on retrieval strategies, intent classification, intermediate representations, constrained decoding, specialist agents, and evaluation methods.

Original post →

More from coding & agent

coding & agent channel →