AI Release Metadata (ai-release-metadata)
A lightweight SDK for standardizing and propagating operational metadata for AI-powered services through existing observability systems.
The Problem
When a production issue occurs in an AI-powered service (e.g., regressions, prompt injection, or unexpected model behaviour), it is often extremely difficult to quickly identify:
- Which prompt version generated the response?
- Which model was used?
- Which Git commit or deployment version is responsible?
- What experiments were active?
While existing observability platforms effectively capture execution metrics such as latency and token usage, manually correlating these spans back to the exact release configuration often results in fragmented and inconsistent metadata.
The Solution
This SDK provides a minimal set of context managers and decorators that automatically enrich existing logs, metrics, and traces with standardized release and environment metadata.
Architecture
flowchart TD
App[Application] -->|with release_context()| Provider[MetadataProvider]
Provider -->|extract| Plugins["Plugins (Git, Env)"]
Plugins --> Context(("Release Context"))
Context -->|get_current_context()| Exporters["Exporters"]
Exporters --> OTEL["OpenTelemetry / Structlog"]
Design Principles
- Minimal developer overhead: Drop-in context managers and decorators.
- Existing Observability First: Integrates directly into your current logging/tracing pipelines instead of requiring a new dashboard.
- A reusable building block: This explores one reusable approach for improving traceability between production behaviour and deployed AI releases.
- Framework agnostic: Works with FastAPI, Django, or raw scripts.
- Vendor agnostic: Does not wrap or depend on specific LLM provider SDKs.
- Explicit over magic: Developers explicitly define trace boundaries.
- Existing observability first: Propagates metadata to your existing tools (like OpenTelemetry) rather than acting as a standalone backend.
- Async-safe by default: Thread-safe and async-safe context propagation.
Features
- Automatic Metadata Discovery: Detects release and runtime information from the execution environment through a pluggable metadata system.
- Application-level Instrumentation: Instruments business operations instead of wrapping LLM SDKs, allowing the library to remain provider-agnostic.
- Integrations: Natively exports into
structlog(JSON logs) and OpenTelemetry. - Async First: Safe to use in high-throughput
asyncioapplications such as FastAPI.
Use Cases
1. Diagnosing Latency Regressions by Git Commit
If latency spikes on a specific LLM feature, you can query Datadog/Grafana for all generate traces grouped by ai.git_sha. The SDK ensures this metadata is accurately attached to every span, instantly identifying the deployment that caused the regression.
2. A/B Testing Prompt Templates
When rolling out a new prompt (e.g. prompt_version="v2.1"), you can seamlessly track the token usage, error rates, and user feedback specifically for that variant.
Documentation
- Architecture: How context propagation works under the hood.
- Design Decisions: The rationale behind the core architectural choices.
- Metadata Schema: The full list of supported fields.
- Plugin Development: How to build custom discovery sources.
- Exporters: How to export data to OpenTelemetry, structlog, and more.
- Roadmap: Future development ideas and planned integrations.
Installation
pip install ai-release-metadata
To enable the OpenTelemetry integration automatically, install the otel extra:
pip install "ai-release-metadata[otel]"
Usage
Configure the SDK once at application startup using the MetadataProvider.
Plugin Precedence Rules: The SDK resolves metadata conflicts based on the order plugins are provided. The last plugin evaluated takes precedence over the preceding ones. We recommend putting local plugins first, and explicit environment plugins last.
from ai_release_metadata import MetadataProvider
from ai_release_metadata.plugins import EnvPlugin, GitPlugin, GitHubActionsPlugin
# 1. GitPlugin tries to guess from local filesystem
# 2. GitHubActionsPlugin overwrites with CI variables
# 3. EnvPlugin overwrites with explicit user-defined env vars
MetadataProvider(plugins=[GitPlugin(), GitHubActionsPlugin(), EnvPlugin()])
Use the context manager (release_context) or decorator (@capture_generation) within the business logic:
from ai_release_metadata import release_context, get_current_context
async def generate_response(user_id: str, query: str):
# Start a release context block
with release_context(feature="customer-support", model="gpt-4o", prompt_version="v2.1"):
# ... LLM calling logic ...
# Dynamically append runtime information to the context
ctx = get_current_context()
if ctx:
ctx.add_document("doc_1")
ctx.add_tag("user_tier", "premium")
All logs emitted within the with release_context(...) block will automatically have the metadata appended under the ai key.
OpenTelemetry Integration
If you use OpenTelemetry, you can automatically enrich the currently active span with the release context. It gracefully ignores spans if none are active.
from ai_release_metadata.integrations.opentelemetry import enrich_span
from opentelemetry import trace
tracer = trace.get_tracer(__name__)
with tracer.start_as_current_span("generate_response") as span:
with release_context(model="gpt-4o"):
# Enrich the span with git_sha, environment, model, etc.
enrich_span()
# ... LLM calling logic ...
Running the Demo
This repository includes a FastAPI demo simulating a generation endpoint.
- Activate a virtual environment and install dependencies:
python3 -m venv venv
source venv/bin/activate
pip install structlog fastapi uvicorn pydantic
- Run the application:
export GIT_COMMIT=a1b2c3d
export ENVIRONMENT=local
export LOG_FORMAT=console
PYTHONPATH=. uvicorn demo.app:app
- Send a request to observe the enriched telemetry in the terminal:
curl -X POST http://127.0.0.1:8000/generate \
-H "Content-Type: application/json" \
-d '{"model": "gpt-4o", "prompt_version": "v2.0", "user_query": "I cannot login"}'
Metadata
Release files for ai-release-metadata 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| ai_release_metadata-0.1.0.tar.gz | 14.4 kB | Without provenance | Details for ai_release_metadata-0.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| ai_release_metadata-0.1.0-py3-none-any.whl | Python 3 | none | any | Details for ai_release_metadata-0.1.0-py3-none-any.whl |
Total release size: 25.8 kB
Release files / ai_release_metadata-0.1.0.tar.gz
| Download URL | ai_release_metadata-0.1.0.tar.gz |
|---|---|
| Size | 14.4 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
10b3418d74e095eb5dd3b104fa9f419afa0d855c1f13d41123504f3f662baf3f
|
|
BLAKE2b-256 checksum How to use checksums |
a1f808b7a3bf3576539a63814fd2e108e179d760404d5ffb478a25546ca97bb2
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.4
|
Release files / ai_release_metadata-0.1.0-py3-none-any.whl
| Download URL | ai_release_metadata-0.1.0-py3-none-any.whl |
|---|---|
| Size | 11.4 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
fc71d0469e76c2e2ce34b07c1e222321d4d3cbfb23cee8a17be32a9869910572
|
|
BLAKE2b-256 checksum How to use checksums |
6677287891ab6bcf3ffda6ae99168f16108c62114f97d7d6c4f1246e61eaddef
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.4
|
Security for ai-release-metadata 0.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.