---
title: "Demo 12 — `ai-classifier` plugin"
---

# Demo 12 — `ai-classifier` plugin

**Module brief:** [release notes, Module 12](/docs/proxy/releases/release-notes-v040/)

## What it does

> Tag LLM-generated SQL at the proxy via `application_name`
> keywords + generated-by markers. Downstream policy plugins
> (token-budget, llm-guardrail) read the tag.

## Use cases

- **Cost attribution.** "How much DB load is coming from agent
  traffic vs human traffic?" — tag-based query metrics.
- **Compliance.** Auditors need to prove non-deterministic LLM
  output didn't escape the schema; the tag lets you slice the
  audit log by traffic source.

## What this demo shows

Three connection styles, three classification outcomes:

| application_name | classified | agent_id | model_id |
|---|---|---|---|
| `psql` | no | — | — |
| `gpt-shopper` | yes | `gpt-shopper` | — |
| `ClaudeAgent-claude-opus-4-7` | yes | `ClaudeAgent-claude-opus-4-7` | `claude-opus` |

Plus a SQL with a `/* generated by GPT-4 */` marker → also
classified, model guess `gpt-4`.

## Run it

```bash
cd demos/v0.4.0/12-ai-classifier
./demo.sh
```

Walks through the four cases, prints the per-`request_id` KV keys
the classifier wrote (visible via `/admin/kv/ai-classifier/`).

## Implementation pointer

`HeliosDB-Proxy-Plugins/ai-classifier/src/lib.rs`. Pure-function
classifier in `classify`; KV writes in `pre_query`. Six unit tests
cover each detection path.

## HeliosDB compatibility

Backend-agnostic.