Quickstart ⚡¶

Get EdgeSLM running locally on your workstation or server in under 2 minutes.

1. Installation¶

Dependencies are managed with uv and pinned in uv.lock. On Windows, it automatically installs PyTorch with CUDA 12.4 support; on Linux, it installs the lightweight CPU build.

git clone https://github.com/Prajwal7-2002/EdgeSLM.git
cd EdgeSLM
uv sync

Run Unit Tests¶

Run the complete test suite to verify extraction, chunking, policy scanning, and model abstractions:

uv run pytest
# Output: 51 passed in ~8s (Zero GPU or model weights required)

2. Start the Local Privacy API¶

Launch the local zero-data-leakage inference server:

uv run uvicorn src.api.app:app --port 8000

Once running: - Interactive UI: Open http://127.0.0.1:8000/docs in your browser. Drag and drop any .pdf, .docx, .eml, .txt, .md, or .env file directly into /classify-file for an immediate security verdict. - Single Prompt Classification:

curl -X POST http://127.0.0.1:8000/classify \
  -H "Content-Type: application/json" \
  -d '{"text": "Use svc_ingest / Copper!2291 for the ledger service until Monday."}'

Example JSON Response¶

{
  "label": "RESTRICTED",
  "confidence": 1.0,
  "policy_override": true,
  "override_reason": "high-entropy credential assignment",
  "scores": {
    "PUBLIC": 0.0,
    "INTERNAL": 0.0,
    "CONFIDENTIAL": 0.0,
    "RESTRICTED": 1.0
  },
  "chunks": 1,
  "latency_ms": 0.85
}

3. Scanning Folders & Directories¶

Scan an entire directory of mixed documents recursively and export audit results to CSV:

uv run python src/scan.py path/to/folder --csv results/scan.csv

/classify-file and src/scan.py support: - PDF documents (.pdf with extractable text layers) - Word documents (.docx) - Email archives (.eml with RFC 822 header parsing) - Configuration & Code (.env, .yaml, .json, .py, .sql)

4. Evaluating on Real Documents (`real-bench-100`)¶

To run the balanced 100-document real benchmark (SEC 8-K filings, Enron emails, GitLab SOPs, and real cryptographic keys) without manual spreadsheet labeling:

uv run python src/data/build_real_bench_100.py