Repository navigation
The original Irminsul LLMOps stack, kept as-is before the v1 rework.
What v0 is
- Llama 3.1 8B fine-tuned with QLoRA (52K Alpaca examples, 3 MLflow experiments)
- FastAPI RAG server: LangChain RetrievalQA over a Pinecone index (MiniLM embeddings)
- Groq-hosted Llama 3.3 70B for the Hugging Face Spaces demo, with a Tavily web fallback
- Prompt-injection and domain-relevance guardrails, typed API responses, health endpoint
- Docker image and Azure Container Apps deploy script
Known limitations (fixed in v1)
- Weekly corpus ingest wasn't idempotent: random vector IDs duplicated the corpus each run
- No release or leak filtering on scraped wiki content
- Sources shown as filenames, not citations; no automated evaluation
v1 is being built on the v1 branch: official-source corpus with patch versions, hybrid retrieval with re-ranking, citations, a leak filter and an eval suite.