Support Agent Kit

Answer tickets from your docs with citations, auto-send only when confident, escalate the rest. $197, one-time. Python source, MIT with a no-resale-as-template clause.

Triage

Category and priority per ticket, with a keyword fallback if the model output is malformed.

RAG + citations

pgvector store (HNSW, upsert ingest) and an in-memory fallback. Every answer cites retrieved doc ids; invented citations are rejected.

Confidence gate

Auto-send above your threshold; below it, urgent, uncited or unparseable, the ticket escalates to a human with a reason.

Evals

Golden set reports the wrong-answer rate among auto-sent replies, plus misses. Usable as a CI gate.

Standalone

Vendors the agent_forge core runtime. Tests run offline with a mock LLM.

Guides

ARCHITECTURE, PGVECTOR, EVALS, SAFETY and runnable examples. Synthetic data only.

Quickstart

pip install -e ".[dev]"
pytest
python examples/quickstart.py
support-kit evals

Honest limits

The offline mock LLM is extractive; real quality depends on your model, docs and threshold.
The bundled embedding is a hashed bag-of-words for tests; use a real embedding model in production.
The gate reduces wrong answers, it does not eliminate them.

*MIT with a clause forbidding resale of the source as a template/boilerplate/course. See LICENSE. Try the interactive threshold demo.