TENGWAR - Private RAG Assistant
100% private enterprise RAG platform (Retrieval-Augmented Generation) for teams that need AI-powered document search without sending data to the cloud.
Problem & Context
Many SMBs require internal knowledge assistants but can't send proprietary documents to external LLM APIs (legal, GDPR, trade secrets, IP protection). TENGWAR emerged to solve this: a plug-and-play hardware + software package running locally, embedding company documents (PDFs, Word, Excel, PowerPoint, confluence exports, etc.) into vector databases and providing ChatGPT-like Q&A interfaces—with precise source citations—without any data leaving the premises. Supports 100+ languages, multi-user access control and integrates with internal authentication systems.
Responsibilities
- Full-stack development: UI, backend, AI pipeline & hardware selection
- RAG architecture: embedding models, vector databases (pgvector), retrieval strategies
- Multi-format ingestion: PDF, DOCX, XLSX, PPTX, TXT, MD, HTML (20+ formats)
- Real-time chat interface: Next.js UI with SignalR streaming for progressive responses
- Source citation overlay: every answer links to exact document & page number
- Role-based access control: departments & user permissions on document collections
- Hardware integration: Linux-based appliance configuration (Debian, systemd services)
- Multi-language translation layer: automatic query & response translation (100+ languages)
- Deployment & support: client onboarding, training & iterative improvements
Architecture & Stack
- Frontend: Next.js 14 (React, TypeScript), Tailwind CSS, SignalR client
- Backend: ASP.NET Core 8 (C#), SignalR WebSocket streaming
- RAG Pipeline: Python (LangChain, sentence-transformers), pgvector (PostgreSQL extension)
- Embedding models: multilingual sentence-transformers (local inference)
- LLM inference: Ollama (local models: Llama, Mistral, etc.) or external API fallback
- Document ingestion: Apache Tika + custom parsers for metadata extraction
- Database: PostgreSQL 16 + pgvector for vector search
- Authentication: LDAP / Active Directory integration + JWT tokens
- Hardware: Intel NUC / custom Linux appliance with GPU (NVIDIA RTX for embeddings)
- Deployment: Docker Compose orchestration, systemd service management
- Monitoring: Prometheus + Grafana for usage analytics
Outcomes
- Enabled 100% private AI assistant for clients unable to use cloud LLMs
- 20+ file format support eliminated manual document conversion workflows
- Precise source citations increased user trust & compliance auditability
- Role-based access ensured departmental data isolation & security compliance
- Multi-language support expanded addressable market to non-English organizations
- Plug-and-play hardware model reduced deployment friction & IT burden
Learn More
Official product page: tengwar.net ↗