Open-source stack
Open, proven, swappable.
We build custom knowledge and retrieval systems from open-source components, integrated with what you already run. We don’t commit to a stack in advance; the mapping step decides. The code is yours.
- Approach
- Open source first
- Chosen by
- The mapping step, on evidence
- Code
- Yours from the first commit
- Licences
- Checked before you commit
In short
An open-source RAG (retrieval-augmented generation) stack combines open tools for document ingestion, vector search, model serving and evaluation to build an AI assistant that answers from your own documents. DeepVox builds these systems from components such as Onyx, Haystack, LlamaIndex, Qdrant, pgvector, vLLM and LiteLLM, chosen per client, and hands over the source code.
public/videos/platform/open-source-stack.mp4 Then run python scripts/video-posters.py for the cover image. The portfolio we go deep in
Every layer of a knowledge system.
| Layer | Tools |
|---|---|
| Knowledge management | BookStack, XWiki |
| Assistant and retrieval | Onyx, Haystack, LlamaIndex, LibreChat, or the DeepVox platform |
| Document ingestion | Docling, Unstructured, Apache Tika, Tesseract |
| Index and vector store | PostgreSQL with pgvector, Qdrant, OpenSearch |
| Embeddings and reranking | BGE-M3, bge-reranker |
| Permissions and identity | Open Policy Agent, Keycloak, Entra ID, purpose-built services |
| Models | Qwen, Mistral and other open-weight models; hosted providers where you prefer |
| Model gateway | LiteLLM |
| Model serving | vLLM, SGLang, Ollama |
| Voice | Whisper and open speech recognition, open text-to-speech |
| Observability and evaluation | Langfuse, Phoenix, Grafana stack |
| Workflow and orchestration | n8n, custom services, Kubernetes, Helm |
Why open source
Ownership you can audit.
Code you own
Source for everything we build, from the first commit, in open formats with documented APIs.
Nothing hidden
Your security team can read every component that touches your data.
Swappable parts
Each layer can be replaced as better tools appear, without starting again.
No licence forfeiture
Exit cost is knowledge transfer: runbooks and a rehearsed exit test.
Licence literacy included
“Open” doesn’t always mean open.
Products relicense, and many “open” platforms put the SSO, role-based access and audit features regulated teams need behind a paid tier. We check licences for every component and flag it before you commit.
Where our own DeepVox platform is a better fit for the assistant layer, we say so, and score it against the open-source options on the same criteria.
FAQ
Questions, answered.
What is a RAG stack?
Retrieval-augmented generation: documents are ingested and indexed, the best passages for each question are retrieved, and a language model writes an answer from them with citations. A stack is the set of tools doing each of those steps.
Do you always use the same tools?
No. The mapping step chooses components against your requirements, existing infrastructure and licences. The list above is the portfolio we know deeply.
Do we get the source code?
Yes, for everything we build, from the first commit, when self-hosted.
Open source or the DeepVox platform?
Open source for full ownership and control; the DeepVox platform for speed and a supported product. We score both for you.
Related
Custom knowledge & retrieval
What we build with this stack.
Learn more →Build vs buy
Open source, platform or SaaS.
Learn more →Models of your choice
Open-weight, fine-tuned or hosted.
Learn more →Start with two weeks of evidence, not a sales call.
A fixed-price discovery sprint, credited against whatever comes next. Or write to sales@deepvox.ai.