Skip to main content

HB-Eval SDK

The official Python SDK for HB-Eval OS — the reliability operating system for agentic AI. Evaluate any agent trajectory against five reliability metrics and receive a tier certification, in a few lines of code.

Install

pip install hb-eval-sdk

For the LangChain integration:

pip install hb-eval-sdk[langchain]

Quick start

from hb_eval_sdk import HBEvalClient

client = HBEvalClient(
    api_key="...",          # identifies your project
    aes_key="...",          # encrypts your payload (base64, 32 bytes)
    signing_secret="...",   # signs your request (base64; never transmitted)
)

result = client.evaluate({
    "trajectory": [
        {"step": 1, "action": "chain_start"},
        {"step": 2, "action": "tool_call", "tool": "search"},
        {"step": 3, "action": "chain_end"},
    ],
    "sub_tasks": 3,
    "constraint_violations": 0,
    "recovery_episodes": [],
    "agent_id": "my-agent",
})

print(result.verdict, result.tier)
print(result.metrics)   # pei, irs, frr, ti, csi

The five metrics

Every evaluation returns five reliability metrics. Any of them may be None when it is genuinely undefined for a given run, and None always means "not measured" — never "scored zero".

  • PEI — Planning Efficiency Index
  • IRS — Intentional Recovery Score (None when the run had no faults)
  • FRR — Failure Resilience Rate
  • TI — Traceability Index (None when no judge evaluation was made)
  • CSI — Consistency Stability Index (None without enough history)

LangChain

from hb_eval_sdk import HBEvalCallback

callback = HBEvalCallback(api_key="...", aes_key="...", signing_secret="...")
agent.run(task, callbacks=[callback])
print(callback.last_result.verdict)

The callback observes the real run — counting genuine tool errors and detecting actual fault-and-recovery patterns — rather than assuming a clean execution.

Credentials

Your project has three credentials, issued together when the project is created. The API key is sent on each request to identify you. The AES key encrypts your payload locally. The signing secret signs your request and is never transmitted — it proves the request genuinely came from you, even to an observer who has seen your API key.

Links

License

MIT

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

hb_eval_sdk-2.6.0.tar.gz (42.8 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

hb_eval_sdk-2.6.0-py3-none-any.whl (45.3 kB view details)

Uploaded Python 3

File details

Details for the file hb_eval_sdk-2.6.0.tar.gz.

File metadata

  • Download URL: hb_eval_sdk-2.6.0.tar.gz
  • Upload date:
  • Size: 42.8 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.14.5

File hashes

Hashes for hb_eval_sdk-2.6.0.tar.gz
Algorithm Hash digest
SHA256 35491db7e6294e5471284a7d19b079585ec368572053550b6cd005c714d3b6bc
MD5 c0eb7ce8af8c1ec0ae1a815d054cbb15
BLAKE2b-256 44e31a0af1c4bb189a8bd506df1c6b7b1f9406c6a3ad860a57219023b03fa79e

See more details on using hashes here.

File details

Details for the file hb_eval_sdk-2.6.0-py3-none-any.whl.

File metadata

  • Download URL: hb_eval_sdk-2.6.0-py3-none-any.whl
  • Upload date:
  • Size: 45.3 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.14.5

File hashes

Hashes for hb_eval_sdk-2.6.0-py3-none-any.whl
Algorithm Hash digest
SHA256 831e5dd382c8c061884eb959c315c72dec30cbc960be14f3f12800d101c819fe
MD5 56b4ecb7484ff37f3676c73582c73ab0
BLAKE2b-256 1b9ff2ae8a37f9141c1eafa270dbf433a69fd77bef9f3b2558b239ff27640ba9

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page