01 · retrieve
The optional local stack uses dense + sparse retrieval (BGE + BM25, RRF fusion). This sample proxy does not call that stack.
← Incident workspace · Sample model console; not retrieval-grounded
OpsLM is Qwen3-4B fine-tuned (QLoRA) on a decontaminated DevOps/MLOps corpus, wrapped in a hybrid-retrieval stack and gated by an eval-first platform. Every claim behind it traces to a measured number in the historical research reports. This console is a separate sample experience; a matched SFT quality improvement is not established.
→ See the measured inference benchmark — vLLM vs Ollama on a real Tesla T4 under a 1→16 concurrency sweep: 13.4× throughput scaling, a prefix-cache control run, and the raw JSON committed next to every chart.
— OpsLM v1 · Qwen3-4B (QLoRA) · sample console
opslm:~$ Ask me about Kubernetes, Docker, Terraform, or MLflow. This console shows sample responses unless a model endpoint is configured. The endpoint proxy does not perform retrieval.
An endpoint is configured; availability has not been checked. This proxy does not attach retrieved context.
The optional local stack uses dense + sparse retrieval (BGE + BM25, RRF fusion). This sample proxy does not call that stack.
The local ChatService can attach context to generation. OpsLM SFT used 534 training pairs; this does not establish a task-quality gain.
Prompt-injection quarantine at ingest and secret redaction — a poisoned doc contributes zero chunks.
CI checks 15 committed-report assertions and two live retrieval thresholds. It does not rerun the historical model-quality studies.
Qwen3-4B base · QLoRA (4-bit NF4) · merged 16-bit + GGUF Q4_K_M
SFT on a 593-pair split (534 train). DPO code exists; a completed DPO run and matched SFT quality gain are not established.
OpenAI-compatible endpoint · Ollama / vLLM / llama.cpp
speculative decoding · guided JSON · quant frontier · TPOT