assay-it 0.8.0rc1

High-performance evaluation framework for LLM agents

Assay Python SDK

Record deterministic traces from your Python agents for regression gating.

🚀 Golden Quickstart

The fastest way to regression test your AI agent.

1. Installation

pip install assay

2. Record (record.py)

Run your agent through the SDK to capture a trace. Pass your tool functions to tool_executors so Assay can record their inputs and outputs.

import os
import openai
from assay_sdk import TraceWriter, record_chat_completions_with_tools

# 1. Setup Client & Tools
client = openai.OpenAI(api_key=os.environ.get("OPENAI_API_KEY", "mock"))
TOOLS = [{
    "type": "function",
    "function": {
        "name": "GetWeather",
        "parameters": {"type": "object", "properties": {"location": {"type": "string"}}}
    }
}]

# 2. Define Execution Logic (The "Real" Code)
def get_weather(args):
    return {"temp": 22, "location": args.get("location")}

# 3. Record the Loop
writer = TraceWriter("traces/quickstart.jsonl")
result = record_chat_completions_with_tools(
    writer=writer,
    client=client,
    model="gpt-4o",
    messages=[{"role": "user", "content": "Weather in Tokyo?"}],
    tools=TOOLS,
    tool_executors={"GetWeather": get_weather}, # Link schema -> function
    episode_id="weather_demo",
    test_id="weather_check"
)
print(f"Agent Final Answer: {result['content']}")

3. Configure (assay.yaml)

Tell Assay what to check.

version: 1
model: "trace"
tests:
  - id: weather_check
    input:
      prompt: "Weather in Tokyo?" # Matches the recorded prompt
    expected:
      type: regex_match
      pattern: ".*" # Pass if any content returned (baseline check)

4. Verify

Run the regression gate. This replays your trace against the recorded tool outputs to ensure determinism.

# Verify strictly (fails if any tool call arg changed even slightly)
assay ci --config assay.yaml --trace-file traces/quickstart.jsonl --replay-strict --db :memory:

---

🌊 Advanced: Streaming support

Capture streaming responses while maintaining tool call execution.

from assay_sdk import record_chat_completions_stream_with_tools

# ... setup client & writer ...

result = record_chat_completions_stream_with_tools(
    writer=writer,
    # ... args ...
    stream=True # SDK handles chunk aggregation automatically
    # tool_executors={...} # Required if tools are used
)

Note: The hybrid wrapper (record_chat_completions_stream_with_tools) streams the thinking tokens to the user, executes tools, and then performs a standard follow-up call.

🛡️ Advanced: Privacy & Redaction

Protect sensitive data (PII, API keys) from ever hitting the trace file.

from assay_sdk import TraceWriter, make_redactor

# Create a redactor that scrubs keys and regex patterns
redactor = make_redactor(
    key_denylist={"authorization", "password", "api_key"},
    patterns=[r"sk-[a-zA-Z0-9]{20,}"] # Mask OpenAI keys
)

# Attach to writer - happens automatically on write
writer = TraceWriter("traces/secure.jsonl", redact_fn=redactor)

⚡ Async Support

Native async support for high-throughput applications (FastAPI, etc.) is available via the assay_sdk.async_openai submodule. It provides full parity with the sync API, including loop and streaming support.

❓ Troubleshooting

E_TRACE_EPISODE_MISSING

Cause: The test_id or episode_id in your trace doesn't match what assay ci expected from its config (or implicit default). Fix: Ensure your assay.yaml test IDs match the test_id passed to record_chat_completions....

"Duplicate prompt in strict replay"

Cause: You ran record.py twice without cleaning the trace file, so it contains two identical episodes. assay ci in strict mode doesn't know which one to replay. Fix:

1. Truncate the file before recording: trace_path = "traces/my_trace.jsonl"; open(trace_path, 'w').close().
2. Use unique episode_ids (e.g. UUIDs) for every run.