ContextOps

Cache-aware prompt structure optimizer + LLM-as-judge eval + local cost/usage logger.

Stop paying for the same tokens twice. ContextOps reorders your prompt sections so stable content (system prompt, tools) sits at the top — and variable content (query, history) sits at the bottom — maximizing cache hit rate on Anthropic / OpenAI / DeepSeek / any provider that does prefix caching.

No cloud, no SaaS, no SDK lock-in. Just pip install contextops-tool and go.

---

⚡ Quickstart

pip install contextops-tool

from contextops import optimize, Prompt

p = Prompt(
    query="What's the weather in Berlin?",
    history=[{"role": "user", "content": "Hi!"}],
    documents="Berlin weather API docs...",
    system="You are a helpful weather assistant.",
    tools='[{"name": "get_weather"}]',
    model="gpt-4o",
)

result = optimize(p)
print(result.diff())                            # history → documents → ... → query
print(f"Cache hit: {result.estimated_cache_hit_rate:.1%}")
print(f"Saves ~${result.estimated_cost_savings_usd:.4f} per 1k calls")

Output:

Section order: history → documents → ... → system → ... → query
Cache hit: 71.0%
Saves ~$0.1006 per 1k calls

That's it. Same prompt, same tokens, ~70% cache hit rate instead of ~5%.

---

🤔 Why?

LLM providers (Anthropic, OpenAI, DeepSeek, Google) cache the prefix of your prompt. If the prefix is stable across calls, you pay 10% of the cached-token price instead of the full price.

The trick: keep the prefix stable by putting variable content (query, history) at the end.

ContextOps knows the canonical ordering by stability:

system → tools → role → context → documents → history → query
  ↑ stable                                                ↑ variable

Estimated impact on a typical workload:

Setup	Cache hit rate	Effective $/1M input
Random order	~5%	$X (full price)
ContextOps optimized	~78%	~$0.3·X (10% on cached prefix)

---

🧰 What's in the box

Feature	Description
Cache-aware reordering	Moves stable sections to the top, variable to the bottom. Same total tokens, much higher cache hit rate.
Token counting	tiktoken-based, model-aware (gpt-4o, claude-*, qwen*, fallback to cl100k_base).
Cost estimation	Per-model pricing baked in; estimates $/1k calls before vs after reorder.
LLM-as-judge eval	Built-in metrics: faithfulness, relevance, completeness, conciseness.
A/B testing	Run two prompts over a golden dataset, get structural + quality deltas.
Local SQLite logger	Every LLM call goes to ~/.contextops/calls.db. Zero cloud.
Dataset loaders	.json, .jsonl, .csv golden QA datasets.
Rich CLI	optimize / stats / recent / compare / eval / reset with tables and progress bars.
LiteLLM auto-log (opt)	One line to auto-log every litellm call. pip install "contextops[integrations]"
Bench harness	1000+ prompts through Ollama, LM Studio, OpenRouter, or direct APIs (Anthropic / OpenAI / Gemini / OpenCode-ZEN).

---

📦 Install

# Core (optimizer + logger + eval + CLI)
pip install contextops-tool

# With LiteLLM auto-callback for real LLM logging
pip install "contextops-tool[integrations]"

# With dev tooling (pytest, ruff, mypy)
pip install "contextops-tool[dev]"

# Everything
pip install "contextops-tool[all]"

From source:

git clone https://github.com/QuickLeopard/contextops.git
cd contextops
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev,integrations]"
pytest                                          # 53 tests
python -m contextops_bench smoke               # offline smoke

Requires Python 3.10+.

macOS / Linux gotcha: python not found

If you see zsh: command not found: python, use python3 instead. Or install Python 3.12 via Homebrew and add it to PATH:

brew install python@3.12
export PATH="/opt/homebrew/opt/python@3.12/bin:$PATH"
python --version   # should print 3.12.x

If you want a fully automated setup, run the bootstrap script after cloning:

git clone https://github.com/QuickLeopard/contextops.git
cd contextops
./scripts/bootstrap.sh   # installs Python, creates venv, runs tests + smoke

No python -m venv?

Some Linux distros ship Python without venv:

# Debian / Ubuntu
sudo apt install python3.12-venv

# Fedora
sudo dnf install python3.12-venv

# Or use virtualenv instead
pip install virtualenv
virtualenv .venv

---

📖 Usage

1. Optimize a prompt (Python)

from contextops import optimize, Prompt

p = Prompt(
    query="What's the weather in Berlin?",
    documents="API docs...",
    system="You are a helpful weather assistant.",
    tools="[]",
    model="gpt-4o",
)

result = optimize(p)
print(result.diff())                  # before → after
print(result.optimized_sections)      # [(Section, content), ...]

2. Compare two prompts (Python)

from contextops.eval import compare

report = compare(baseline=bad_prompt, optimized=good_prompt)
print(report["delta"])
# {"tokens": 0, "cache_hit_rate": 0.65, "cost_savings_per_1k_usd": 4.21}

3. A/B eval with LLM-as-judge

from contextops import evaluate_ab, load_dataset, Prompt, LiteLLMJudge

dataset = load_dataset("evals/sample_dataset.jsonl")

baseline = Prompt(system="...", query="", documents="{ctx}", model="gpt-4o-mini")
optimized = Prompt(system="...", documents="{ctx}", query="", model="gpt-4o-mini")

def my_llm(prompt_str: str) -> str:
    return call_my_llm(prompt_str)

report = evaluate_ab(
    baseline, optimized,
    run_fn=my_llm,
    dataset=dataset,
    metrics=["faithfulness", "relevance", "completeness"],
    judge=LiteLLMJudge(),
    on_render=lambda p, item: p.system.replace("{ctx}", item.context),
)

print(report["structural"])   # tokens / cache / cost deltas
print(report["quality"])      # per-metric judge deltas

4. CLI

# Optimize a prompt inline
contextops optimize \
    --system "You are a weather assistant." \
    --query "What's the weather in Berlin?" \
    --documents "API docs..." \
    --model gpt-4o

# Load a prompt from a JSON file
contextops optimize --from-json my_prompt.json

# Side-by-side comparison
contextops compare baseline.json optimized.json

# A/B eval with offline echo judge
contextops eval \
    --baseline evals/baseline_prompt.json \
    --optimized evals/optimized_prompt.json \
    --dataset evals/sample_dataset.jsonl \
    --metrics relevance,completeness,faithfulness \
    --echo --run-fn echo \
    --output report.json

# Real LLM-as-judge
pip install "contextops[integrations]"
contextops eval \
    --baseline evals/baseline_prompt.json \
    --dataset evals/sample_dataset.jsonl \
    --judge-model gpt-4o-mini \
    --metrics relevance,completeness,faithfulness,conciseness

# Local call stats
contextops stats
contextops recent --limit 50

# Reset the local database
contextops reset

5. Auto-log every LiteLLM call

from contextops.integrations import install_callback
install_callback()

import litellm
litellm.completion(model="gpt-4o", messages=[{"role": "user", "content": "hi"}])
# → automatically logged to ~/.contextops/calls.db

---

🆚 Comparison

Tool	What it does	Where ContextOps is different
DSPy	Auto-rewrites prompt text using a dataset	We reorder sections — no dataset, no model rewrite
RAGAS / DeepEval	Evaluate answer quality via LLM-judge	We measure structure + cost, complementary not competing
Langfuse	Cloud LLM observability	We stay local-first: SQLite, no signup
prompt-cache / token-optimizer	Cache responses, compress tokens	We focus on provider cache (Anthropic / OpenAI), not response cache
vaibkumr/prompt-optimizer	Compress text (LLMLingua-style)	We reorder, never change tokens or text

---

🧪 Bench harness

1000+ prompts through Ollama, LM Studio, OpenRouter, or direct APIs:

# Smoke (10 prompts, <30s, no LLM, for CI)
python -m contextops_bench smoke

# Local (100 prompts via Ollama)
python -m contextops_bench local --provider ollama --model llama3.1:8b --n 100

# Cloud via OpenRouter, multi-model, parallel
export OPENROUTER_API_KEY=sk-or-v1-...
python -m contextops_bench cloud --provider openrouter \
    --model openai/gpt-4o-mini,anthropic/claude-3.5-haiku,meta-llama/llama-3.1-8b-instruct \
    --n 1000 --parallel 4

# Direct APIs (bypasses OpenRouter translation — definitive cache signal).
# Pick the provider that matches your target's cache mechanics:
export ZEN_API_KEY=...            # or ANTHROPIC_API_KEY / OPENAI_API_KEY / GOOGLE_API_KEY
python -m contextops_bench cloud --provider direct_anthropic \
    --model claude-sonnet-4-6 --preset-agent realistic --n 30   # explicit cache_control markers
python -m contextops_bench cloud --provider direct_openai \
    --model gpt-4o-mini --preset-agent realistic --n 30          # automatic caching, 50% off
python -m contextops_bench cloud --provider direct_google \
    --model google/gemini-2.5-flash --preset-agent realistic --n 30  # implicit caching, 10% off

About --preset-agent: the bench needs a stable system prompt + tool schema + role across all calls for cache hits to be non-zero. --preset-agent realistic pins all three. On cache-bearing providers (OpenRouter + the four direct_*), the bench auto-applies realistic if you don't pass a preset — with a loud warning explaining why. Use --preset-agent none to opt out and use randomized prompts.

Each run writes:

• bench/results/<label>.csv — every observation (prompt_id, model, tokens, cache hit, cost, latency, error, section order)
• bench/results/<label>.summary.json — aggregated stats with optimized vs baseline deltas

Troubleshooting cache reads showing 0? Read docs/POSTMORTEM_realistic_cache.md — it covers the realistic-preset cache key regression, why OpenRouter drops cache_control markers during translation, and why EchoClient (used in unit tests) hides the bug.

See docs/ACCEPTANCE.md for the formal pass criteria.

---

🗺️ Roadmap

• ✅ v0.1 — reorder, token count, SQLite logger, CLI
• ✅ v0.2 — LLM-as-judge eval + A/B testing + dataset loaders
• ✅ v0.3 — realistic-preset cache-key regression fix + direct providers (Anthropic / Zen / OpenAI / Gemini) + CI bench regression gate + safety-net auto-default on cache-bearing providers. See docs/POSTMORTEM_realistic_cache.md.
• 🔜 v0.4 — RAG curator (multi-signal retrieval + strict threshold)
• 🔜 v1.0 — Access-aware context + audit trail (on-prem / enterprise)

---

📚 Documentation

• docs/ACCEPTANCE.md — formal pass/fail criteria
• CHANGELOG.md — version history
• CONTRIBUTING.md — how to contribute
• SECURITY.md — how to report vulnerabilities
• evals/ — sample datasets and prompts

---

🤝 Contributing

PRs welcome. See CONTRIBUTING.md for workflow, conventions, and release process.

Good first contributions:

• New metric in contextops/judge.py (e.g. safety, format_compliance)
• New provider in contextops_bench/clients.py (e.g. vllm, tgi)
• Better pricing tables for non-USD regions
• Translations of docs/ and README.md

---

📜 License

MIT.

---

✨ Credits

Built with:

• pydantic for prompt modeling
• tiktoken for token counting
• click + rich for the CLI
• litellm for the optional auto-callback