← BACK TO AI LAB

AI LAB / LLMs / Streaming

Streaming JSON Repair

Exploring whether partial structured output can be validated and repaired during streaming.

LLMsStreamingStructured Output
EXPERIMENT
EXP-021
STATUS
inconclusive
DATE
Jul 14, 2026
TAGS
LLMs · Streaming · Structured Output

EXPERIMENT / HYPOTHESIS

Incremental parsing can surface useful validated fields before a complete model response arrives.

  1. 01Stream 200 nested JSON responses
  2. 02Parse complete tokens into an incremental tree
  3. 03Measure time-to-first-valid-field and repair accuracy
Earlier useful data
420ms
Repair success
72%
Runs
200

CONCLUSIONIncremental parsing improved perceived speed, but repair behavior was too provider-specific to standardize without a narrower schema subset.

01

Approach

The parser emitted only fields whose syntax and local schema constraints were complete, while buffering uncertain branches.

  • Incremental tokens
  • Field-level validation
  • Final full-object check
02

Why inconclusive

Provider chunk boundaries and partial escaping produced inconsistent repair opportunities. The complexity outweighed the latency benefit for most flows.

  • Useful for long outputs
  • Risky for side effects
  • Needs constrained schemas