01Sovereign knowledge base
Local multi-agent retrieval across documents in any language. Ask in one language, get answers grounded in sources written in another — without a byte leaving your network.
Local-first · Sovereign · Multi-agent
Your knowledge, your models,
your infrastructure
Sovnex is a self-hosted AI workspace built on three pillars: a multilingual sovereign knowledge base run by local agents, multi-agent RFP analysis and bid drafting, and a local Token Factory that puts every model — including your own — behind a single API. Nothing leaves your network.
02RFP analysis & bid response
Drop in a tender. Agents extract every requirement, pull the matching evidence from your knowledge base, and draft a sourced response for your team to review.
03Local Token Factory
One OpenAI-compatible API in front of every model, local or frontier. Route routine agent traffic to your own models and stop paying tokens for it.
Pillar 01
A sovereign knowledge base,
local agents, any language
Documents, embeddings and prompts stay on your own infrastructure. Specialised agents retrieve and reason across them — and language is never a barrier between the question and the answer.
- Runs entirely inside your network
- Deploy on-premise or in your private cloud. Files, embeddings, prompts and inference all stay on hardware you control, so sensitive material never reaches a third-party API.
- Local multi-agent retrieval
- Specialised agents — retrieve, analyse, draft, review — break a complex question into steps and work through it together, instead of squeezing everything into a single prompt.
- Multilingual and cross-lingual
- The workspace speaks six languages, and retrieval crosses them: ask in French and Sovnex grounds the answer in your English or Chinese documents, citing the original passage.
- Every answer traceable to its source
- Each response points back to the exact chunk it came from, down to the page. Chunking is visible and editable, so your team fixes the source instead of arguing with the model.
One workspace over your own documents — knowledge bases, Q&A, chat and search, answering from a local model.
Pillar 02
From tender document to
a sourced bid response
Upload an RFP, a tender pack or a security questionnaire. Sovnex breaks it into requirements, finds the evidence in your knowledge base, and drafts an answer for each one — every claim carrying its citation.
- 1
Extract every requirement
The tender is parsed into an itemised list of requirements and questions, so nothing buried on page 40 is missed.
- 2
Retrieve the evidence
For each requirement, agents search your past proposals, certifications and datasheets, and assemble the passages that actually answer it.
- 3
Draft the answer
A written response is produced per requirement, in the language of the tender, with each statement linked to the source it came from.
- 4
Review and export
Your team reviews, edits and exports. Confidence scores and citations make it obvious which answers need a human eye first.
app.sovnexai.com
RFP Auto-Answer
Knowledge Bases
Agent Runs
S
Sovnex Team
On-Premise
Draft a response to requirement 4.2: "Describe your data encryption at rest and in transit."
All data is encrypted with AES-256 at rest and TLS 1.3 in transit. Sovnex runs entirely inside your network, so customer data never leaves your infrastructure. Key management integrates with your existing KMS/HSM…
Source: Security_Whitepaper.pdf p.7Confidence: 98%
Ask or paste an RFP requirement…
Pillar 03
One API in front of every model,
so your agents cost less to run
Token Factory is the model gateway bundled with Sovnex. Point any OpenAI SDK at it and reach every provider — or your own local models — through a single interface, with budgets, quotas and usage fully under your control.
- OpenAI-compatible endpoints
- Change one base URL and your existing code keeps working. /v1/chat/completions and /v1/embeddings behave exactly as the OpenAI SDK expects, so nothing else in your stack has to change.
- Every provider behind one interface
- OpenAI, Claude, Gemini, DeepSeek and Qwen alongside local vLLM or Ollama deployments. Group them into channels, map model names, and fail over automatically when one goes down.
- Route routine work to local models
- Day-to-day and batch agent traffic runs on your own Ollama or vLLM deployment at zero external token cost. Only the calls that genuinely need a frontier model go out — and that is where most agent spend goes.
- Budgets, quotas and rate limits
- Issue an API token per team or project, then cap its budget, set an expiry, whitelist the models it may call and throttle its request rate — without handing out a single upstream provider key.
- Usage and cost you can audit
- Every call records the model, token counts, latency and cost. Break spending down by token, user or channel to see exactly where the budget went.
python
from openai import OpenAI
client = OpenAI(
base_url="https://sovnex.internal/v1", ← the only change
api_key="sk-sovnex-••••",
)
client.chat.completions.create(
model="qwen2.5-72b", # or any channel you configured
messages=[{"role": "user", …}],
)OpenAIClaudeGeminiDeepSeekQwenvLLMOllama
Same endpoint whether it routes to a local Qwen today or a frontier model tomorrow.
Every token carries its own quota, expiry and model whitelist — and reports what it spent.
About Sovnex AI
Built to stay inside your walls
Sovnex AI builds a sovereign RAG engine for teams that cannot send their documents to someone else's cloud. The engine is open source, deployed on your own hardware, and answers to your rules.
- Apache 2.0
- The engine is open source — audit it, fork it, run it without us.
- Local-first
- Deployed on your own servers or private cloud, with local inference for sensitive data.
- Hosting in your own infrastructure
- The engine runs on servers you own or rent. Your documents and embeddings are never stored in a cloud we operate.