Metadata-Version: 2.4
Name: tracelm
Version: 0.3.0
Summary: Lightweight distributed LLM execution tracer and profiler
Author: Tapesh Chandra Das
License: MIT
Requires-Python: >=3.9
Description-Content-Type: text/markdown
Provides-Extra: fastapi
Requires-Dist: fastapi; extra == "fastapi"
Provides-Extra: requests
Requires-Dist: requests; extra == "requests"
Provides-Extra: otel
Requires-Dist: opentelemetry-sdk; extra == "otel"

# TraceLM
Open Source LLM Execution Tracer and Profiler

## 1. Overview
TraceLM is a lightweight execution tracing and profiling engine for LLM applications. It is designed for LLM pipelines, RAG systems, and multi-step workflows that require infrastructure-level observability. The runtime model uses async-safe `ContextVar` propagation and enforces a single-root DAG structure for trace consistency.

## 2. Core Capabilities
- Hierarchical span tracing
- Synthetic root span model
- Deterministic trace DAG construction
- Critical path detection
- Slowest span identification
- Token and cost aggregation
- Basic anomaly detection
- Regression comparison between traces
- Chrome Trace Format export

## 3. Architecture
TraceLM execution is CLI-driven. The CLI owns trace lifecycle and creates a synthetic root span per run. Decorators only create child spans under the active parent span. `ContextVar` propagation provides async safety for active span context. Before profiling, the trace DAG is validated. Profiling and summary generation occur after execution completes.

Execution flow:
`CLI run -> create trace -> create root span -> execute user code -> finish root span -> validate DAG -> profile -> store -> print summary`

## 4. Installation
```bash
python -m venv venv
# Linux/macOS
source venv/bin/activate
# Windows (PowerShell)
.\venv\Scripts\Activate.ps1
pip install -e .
```

## 5. Usage
Example instrumented code:

```python
from tracelm.decorator import node


@node("step1")
def step1():
    return 1


@node("step2")
def step2(x):
    return x + 1


def main():
    x = step1()
    return step2(x)


main()
```

CLI commands:

Run:
```bash
python -m tracelm.cli.main run test_app.py
```

List:
```bash
python -m tracelm.cli.main list
```

Analyze:
```bash
python -m tracelm.cli.main analyze <trace_id>
```

Compare:
```bash
python -m tracelm.cli.main compare <id1> <id2>
```

Export (Chrome format):
```bash
python -m tracelm.cli.main export <trace_id> --format chrome
```

## 6. Chrome Trace Visualization
- Export generates `trace_<id>.json`
- Open Chrome
- Navigate to `chrome://tracing`
- Load the JSON file
- Inspect the execution timeline

## 7. Limitations (Current Version)
- Single-process tracing only
- No distributed tracing
- No auto-instrumentation
- No UI dashboard
- No OpenTelemetry exporter yet

## 8. Roadmap
- Span duration histograms
- Regression threshold alerts
- OpenTelemetry compatibility
- Distributed trace support
- Web dashboard
- Performance benchmarking suite

## 9. License
MIT License

## 10. Author
Tapesh Chandra Das
