One core that reasons. Three that ship.
VxThinker is a production-grade AI platform: three fine-tuned specialist LLMs — cloud, code, support — coordinated by one reasoning core. Grounded in your private data via FAISS and chain-of-thought. No external API calls. Nothing leaves your network.
Open stack · runs on your hardware
The 3 + 1 architecture
A reasoning core, plus three specialists that do one job well
A universal /v1/ask dispatcher classifies each prompt and forwards it to the right backend. Every specialist gracefully falls back to its base model, so every route works from a clean checkout.
VxThinking
v1.2 · coreRAG + planning + chain-of-thought reasoning. Powers the FAISS-backed retrieval endpoints, ticket/sprint/forecast intelligence, and incident-pattern detection.
- FAISS vector search over your private data (all-MiniLM-L6-v2)
- ReasoningEngine: search → analyze → synthesize → decide
- Routes provisioning intent to the Go agent
VxCloud
v1.0DevOps / IaC / SRE specialist — Terraform, Kubernetes, Helm, Ansible, runbooks, and cost optimization, with safety rules baked in.
- Every security-relevant line is annotated
- K8s manifests always include resources.limits
- IAM generated least-privilege by default
VxCoder
v1.0Code generation, multi-file edits via XML SEARCH/REPLACE diffs, PR review, and test writing (pytest / vitest).
- Multi-file edits as reviewable SEARCH/REPLACE diffs
- Diff review: correctness · security · readability
- Generates pytest / vitest suites
VxSupport
v1.0IT support, docs Q&A, runbook lookup, and Jira-style ticket auto-answer — always structured for action.
- Answers as Diagnosis → Steps → Verify → Escalate
- Cites the sources it retrieved
- Auto-answers tickets from title + body + labels
Grounded, not guessed
Every answer is retrieved from your data, then reasoned over
VxThinking encodes your CSVs, docs, and runbooks into a FAISS index withall-MiniLM-L6-v2. The ReasoningEngine searches, analyzes, synthesizes, and decides — with two cache layers keeping repeat queries fast.
- L1 embeddings cache (TTL 1h) · L2 search cache (TTL 30m)
- GPU auto-detected; all four models share the device
- Provisioning intent returns a Golang payload for the agent
- Fully self-hosted — sovereign and air-gap friendly
Prompt + docs encoded with all-MiniLM-L6-v2 · L1 embeddings cache
One gateway · every model
Ask, generate, search, train — from one FastAPI surface
Hit /v1/ask and let the dispatcher route, or call a specialist directly. Train each model on your own datasets, build the FAISS index, and ship it as a single container on port 8745.
- Universal /v1/ask with force_model override
- LangChain-compatible /generate response shape
- Per-model train.py + auto-train on startup
- Docker compose up — health, stats, metrics built in
1# The universal dispatcher classifies your prompt and forwards it2# to the right specialist — no need to pick a model yourself.3 4curl -s http://llm.vxcloud.io/v1/ask \5 -H 'content-type: application/json' \6 -d '{"prompt":"Write a Terraform module for an S3 bucket with KMS"}' | jq .7 8# Force a specific model when you want fine control9curl -s http://llm.vxcloud.io/v1/ask \10 -H 'content-type: application/json' \11 -d '{"prompt":"Review this diff", "force_model":"codingllm"}' | jq .12 13# Inspect the keyword routing table14curl -s http://llm.vxcloud.io/v1/ask/routes | jq .Let the dispatcher choose
POST /v1/ask classifies the prompt and forwards to thinking / cloud / coding / support — pass force_model to override.
Production-grade by default
Everything a sovereign AI platform should ship with
No data leaves your network
Self-hosted, air-gap friendly. No external API calls — your prompts, datasets, and indexes stay on your hardware.
Universal dispatcher
/v1/ask classifies and forwards to the right specialist; force_model overrides, /v1/ask/routes shows the table.
FAISS-backed RAG
Private vector index with all-MiniLM-L6-v2 and two cache tiers for fast, grounded answers.
Safety rules baked in
VxCloud annotates security-relevant lines, forces K8s resource limits, and generates least-privilege IAM.
Single-container deploy
docker compose up on port 8745. GPU auto-detected; all four models share the device chosen at startup.
Built-in observability
Prometheus + JSON metrics, performance counters, and log inspection under /monitoring/*.
Your data. Your hardware. Your reasoning core.
Bring VxThinker up with one docker compose up, point it at your datasets, and get four cooperating models behind a single API — without a byte leaving your network.