REPOSITORY ANALYSIS

TypeLLM

An open typed-generation layer built on SGLang that leaves autoregressive model architecture and weights unchanged while returning schema-constrained string, integer, number, boolean, or enum Python values.

E2 · source reviewedPublic material or source evidence was reviewed. A fixed-version source review requires checking a pinned revision; this is not a reproduction, performance endorsement, or safety guarantee.
Stars
705
Forks
43
License
Apache-2.0
Last commit
2026-09-26

What Jev does here

Independent fields can run as a batch, while depends_on forms a DAG so later fields can read dependency results and reuse shared prefix caches.

Extends beyond finite choices and booleans to free strings, integers, decimals, image input, optional thinking and permutation averaging for structured-generation workflows such as receipt extraction.

Pinned README, tests and JevBench documentation reviewed on 2026-09-26. Schema constraints guarantee output shape, not semantic correctness or calibrated probabilities; benchmark results are author-reported and were not independently reproduced here.

What task does this project fit?

Extract both bounded decisions and open-valued fields from text or images with an existing SGLang model.

Material analysed:

From input to action

  1. Declare schema types and optional field dependencies.
  2. Batch independent fields or execute the dependency graph.
  3. Return typed values; only eligible boolean/enum fields can include probabilities.

Implementation details that matter

DAG mode is an execution contract

depends_on can reference fields declared later. Dependency values, not their full probability distributions, become later context; explicit batch/sequential modes cannot be combined with this graph mode.

Inspect the supporting source ↗

Permutation averaging has concrete test coverage

test_permutations.py uses fake backends to check option reorderings, reproducible seeds and probability aggregation. Such tests establish orchestration behavior, not model accuracy.

Inspect the supporting source ↗

When to choose it

Consider it when a receipt needs merchant strings and numeric amounts alongside categories, or when one field depends on another. It is not the same model architecture or latency contract as Jev.

How to inspect it yourself

Read the schema restrictions alongside permutation tests. Check that return_probabilities is not requested for unrestricted Text or Numeric fields and that cycles/unknown dependencies are rejected by the chosen version.

Limits of this review

  • No SGLang inference or benchmark was reproduced.
  • Constrained output shape does not guarantee extracted facts or calibrated confidence.
Primary material used in this analysis