driftbase 0.2.2

Behavioral drift monitoring for AI agents.

Driftbase

Behavioral drift detection for AI agents.

Fingerprint your agent's behavior in production. Diff two versions. Get a statistically grounded drift score, financial impact analysis, and plain-English verdict — before your users notice something changed.

---

What it does

Most observability tools tell you your agent ran. They don't tell you it behaved differently than last week.

Driftbase records behavioral signals from each run — tool call sequences, latency percentiles, token cost bloat, and decision outcomes — and computes a drift score when you compare two versions. Every deploy gets a verdict: ship, monitor, review, or block.

• One decorator, zero config — wraps your existing agent entry point.
• Financial impact — translates token bloat directly into Euro (€) cost deltas.
• Data never leaves your machine — local SQLite, zero egress architecture.
• Edge PII Scrubbing — built-in regex engine strips emails, IBANs, and IPs before hashing.
• Plain-English verdicts — tells you exactly what changed and what to do next.

---

The 60-Second Quickstart

See the drift engine in action right now without writing any code.

1. Install

For production tracking (decorator only):

pip install driftbase

For local analysis (CLI + diff engine):

pip install 'driftbase[analyze]'

The base install provides the @track() decorator with minimal dependencies (pydantic, httpx). The [analyze] profile adds numpy, scipy, and rich for statistical drift computation and terminal UI.

2. Run the synthetic demo

This command instantly populates your local database with 50 baseline runs (v1.0) and 50 regressed runs (v2.0 with higher token usage and hallucinated tool calls).

driftbase demo

3. Diff the versions

driftbase diff v1.0 v2.0

You will immediately see a rich terminal UI detailing the financial impact, tool sequence changes, and a root-cause hypothesis:

────────────────────────────────────────────────
  DRIFTBASE  v1.0 → v2.0  ·  50 vs 50 runs
────────────────────────────────────────────────

  Overall drift      0.28  [0.24–0.31, 95% CI]

  Decisions          0.39  · MODERATE
    └─ escalation rate jumped from 5% → 17%
  Latency            0.34  · MODERATE
    └─ p95 increased 4970ms → 6684ms
  Errors             0.03  ✓ STABLE

╭──────────────── Financial Impact ─────────────────╮
│ Cost increased by 223.6%. This change will cost   │
│ an additional €10.46 per 10,000 runs.             │
╰────────────────────────────────────────────────────╯

╭──────────────── VERDICT  ⚠ REVIEW ────────────────╮
│ Most likely cause:                                 │
│   → Tool 'search_knowledge_base' dropped from      │
│     baseline — no longer being called              │
╰────────────────────────────────────────────────────╯

4. Inspect the failure

Grab a run ID from the synthetic data and see exactly what happened. This proves your raw text is safely hashed at the edge:

driftbase inspect <RUN_ID>

---

Instrument Your Code

Once you see the value, drop Driftbase into your actual application. It auto-detects LangChain, LangGraph, LlamaIndex, and raw OpenAI clients.

In your production container (requires only pip install driftbase):

from driftbase import track
import openai

@track(version="v1.0")
def run_agent(prompt: str):
    client = openai.Client()
    return client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}],
        tools=[{"type": "function", "function": {"name": "query_db"}}]
    )

# Run your code normally. Telemetry is silently captured in ~/.driftbase/runs.db
run_agent("What's the revenue for Q4?")

On your laptop or in CI (requires pip install 'driftbase[analyze]'), let the data decide:

driftbase diff v1.0 v2.0 --format md > pr_comment.md

The @track decorator has zero production overhead (pydantic + httpx only). Statistical analysis runs locally with numpy/scipy.

---

EU Compliance & Data Sovereignty

Driftbase is engineered specifically for European teams with strict compliance obligations (GDPR, EU AI Act, DORA).

• Zero-Egress: The free tier runs entirely on your machine. Data never goes to US servers.
• Structural Hashing: We analyze the structure of the behavior, not the sensitive context.
• Edge PII Scrubbing: Enable the scrubber to redact personal data before it even hits your local disk.

# Enable PII redaction
export DRIFTBASE_SCRUB_PII=true

# Configure your custom enterprise rates for accurate cost deltas
export DRIFTBASE_RATE_PROMPT_1M=2.50
export DRIFTBASE_RATE_COMPLETION_1M=10.00

---

Driftbase Pro (Team Sync)

Local SQLite is perfect for individual feature branches. But when Engineer A and Engineer B need to compare baselines, or when you deploy to production, you need a shared source of truth.

To sync your local runs to a secure, EU-hosted centralized dashboard:

export DRIFTBASE_API_KEY="your_pro_key"
driftbase push

Learn more about Driftbase Pro

---

CLI Reference

Command	Description
driftbase demo	Inject synthetic runs to test the engine
driftbase diff <v1> <v2>	Compare two versions (use --format md for CI/CD)
driftbase inspect <id>	Deep-dive into a specific run's execution path
driftbase watch -a <v1>	Live terminal dashboard monitoring incoming runs
driftbase runs -v <v1>	List all local runs for a version
driftbase config	View your current local configuration and DB path
driftbase push	Sync local database to the Driftbase Pro cloud

---

License

Apache License 2.0. See LICENSE for details.