Add prompt-compression example: BANKING77 65K -> 3.2K (95.1%)

[ZhengyaoJiang]

Summary

New examples/prompt-compression/ showing the minimize-chars-with-accuracy-floor shape of Weco optimization, complementing the maximize-accuracy shape in examples/prompt/.

The example deliberately bloats a BANKING77 classifier system prompt to 65,887 characters (mimicking real production prompt patterns: operating principles, per-class blocks, FAQ, worked examples), then has Weco compress it while constraining accuracy to stay at or above the baseline minus 2pp.

Headline result

claude-opus-4-7 × 50 steps: 65,887 → 3,229 chars (95.1% reduction) holding accuracy at 0.7500 on a 200-sample BANKING77 test slice (baseline 0.7700, threshold 0.7500). gpt-5.5 found a different plateau at 6,828 chars (89.6%).

See the trajectory: https://weco.ai/share/XSRQdS7vfMdt9beD3KR1tlhg7By-FFIo

What's in the folder

File	Purpose
optimize.py	Baked baseline + classify(query, model). Weco mutates only the SYSTEM_PROMPT string in the marked WECO-MUTABLE REGION.
eval.py	200 BANKING77 samples → emits accuracy:, chars:, metric:.
labels.py	77 canonical labels + robust parse_predicted_label().
build_bloated_prompt.py	Deterministic baseline generator.
bake_optimize.py	One-shot: writes optimize.py from the generator.
measure_baseline.py	Measures baseline accuracy → baseline_accuracy.json.
prompt_guide.md	--additional-instructions content for the optimizer.
baseline_accuracy.json	Pre-measured 0.7700 (gpt-5-mini, seed=0).
README.md	Setup + run + headline result + structure of the compressed prompt.

Also updates examples/README.md to add this example to the table of contents, the at-a-glance table, and the quick-starts section.

Note on the baseline

The 65K-char baseline is synthetic — generated to mimic real production classifier prompts at the same scale as a representative customer prompt. BANKING77 itself is real and public (PolyAI/banking77 on HuggingFace). The compression ratios match what we've seen on real customer prompts of similar size and shape.

Test plan

• Smoke-tested locally: optimize.py imports, all 77 labels round-trip through parser, eval.py imports and threshold loads from baseline_accuracy.json
• Independent eval of the 3,229-char winner: accuracy=0.7500, parse_rate=1.0
• Reviewer: spot-check weco run end-to-end on a fresh checkout

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 395bf80883

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid uppercasing pin inside unrelated words

The global replacement replace("pin", "PIN") rewrites any pin substring, which corrupts generated prompt text such as topping_up_by_card into topPINg up by card (and the same typo propagates into optimize.py via bake_optimize.py). Because this example’s optimization target is classification accuracy under a strict threshold, injecting malformed intent wording into the baseline prompt can systematically hurt intent recognition for affected classes and distort the measured baseline/threshold used by eval.py.

Useful? React with 👍 / 👎.