docs(examples): add semantix-ai semantic validation example

[labrat-akhona]

Summary

Adds a runnable example to examples/ showing how outlines (structural decoding) and semantix-ai (semantic intent validation via local NLI) compose:

• outlines guarantees the response parses against a JSON schema.
• semantix-ai checks that the generated string actually means what was asked for, using a local cross-encoder NLI model (no API key required).

The script generates a structured customer-support reply with outlines, then scores the reply field for tone with semantix-ai. It also scores a deliberately rude counterexample with the same judge — same schema, different semantics — to make the failure mode outlines cannot catch on its own concrete.

What changed since the previous revision

Per @RobinPicard's earlier feedback (semantix-ai doesn't depend on outlines, so docs/community/examples.md wasn't the right home), this revision:

• Drops the docs/community/examples.md entry.
• Adds a single self-contained worked example at examples/semantix_semantic_validation.py.
• Rebases onto current main so the diff is just the new file.

Run

pip install outlines openai "semantix-ai[nli]"
python examples/semantix_semantic_validation.py

semantix-ai: https://github.com/labrat-akhona/semantix-ai

[RobinPicard]

I took a look at the repo but could not see outlines in the dependencies. How are you using it?

[labrat-akhona]

Hey @RobinPicard — fair catch, thanks for taking the time to look.

You're right: semantix-ai doesn't depend on outlines. It operates at a different layer — outlines constrains generation (schema/grammar), while semantix validates the semantics of an already-generated string against a natural-language intent via local NLI. They're complementary rather than one building on the other.

Given that, I don't think examples.md is the right home for this link — that page is for projects actually using outlines. Two options, happy with either:

1. I close this PR, build a real worked example (outlines for structured decoding → semantix for post-generation intent check) in a notebook, and resubmit once it genuinely uses the library.
2. If you'd rather not carry integration examples for complementary validators at all, I'll just close it.

Let me know which you prefer and I'll action it.

[RobinPicard]

Hi @labrat-akhona, sorry for the delay. 1 sounds good, it could in the examples directory, we already have scripts showing integration with other products there

[labrat-akhona]

Hey @RobinPicard — thanks for the steer. Pushed a fresh revision:

• Dropped the docs/community/examples.md entry.
• Added a single runnable example at examples/semantix_semantic_validation.py.
• Rebased onto current main so the PR diff is just the new file (+94 / -0).

The example uses outlines.Generator with a JsonSchema to structurally constrain a customer-support reply (matching the Generator(model, JsonSchema(...)) pattern from examples/pick_odd_one_out.py), then runs the reply field through a local NLI cross-encoder from semantix-ai to validate the tone. It also scores a hand-written rude counterexample with the same judge to make the schema-valid-but-semantically-wrong failure mode explicit.

Happy to adjust naming / docstring style / move it into a subdirectory if you'd prefer — just let me know.

[RobinPicard]

Looks good. I just changed my mind on the location, could you put it in examples/community please? Also put in the __init__.py of the community directory a little comment on how it contains examples provided by members of the community on how to use outlines with other libraries?