No framework in the middle
Ingestion, chunking, embedding, retrieval and answering are written directly — no LangChain — so every step can be inspected and tuned.
case study · Applied GenAI · 2026
NeuralDesk
A multi-tenant support platform: each business uploads its own documents, and an AI agent answers its customers from them — citing the source — in a chat widget it can embed anywhere.
the pipeline
- Ingestpdf-parse · mammoth · custom chunkingPDF and Word files split into passages
- EmbedOpenAI embeddings → PineconePassages stored as vectors per tenant
- AnswerGroqA reply grounded in the retrieved passages, with citations
- DeliverSocket.io · embeddable widgetLive chat on any website
- My role
- Sole developer
- For
- Personal project · code on GitHub
- Status
- Code on GitHub · code ↗
- Stack
- Next.js · TypeScript · Node.js + Express · MongoDB · Pinecone · OpenAI embeddings · Groq · Socket.io
the problem
Support teams answer the same questions from policies scattered across PDFs and Word files. A general chatbot doesn't know a business's own rules; NeuralDesk answers from the documents each business gives it, and shows where the answer came from.
what I built
- A RAG pipeline built from scratch: PDF and Word ingestion with custom chunking
- OpenAI embeddings stored and searched in Pinecone
- Answers generated with Groq that cite their source documents
- Role-based, multi-tenant access for admins, agents and customers
- Real-time chat over Socket.io and an embeddable support widget
- A visual automation flow builder
decisions that mattered
Every answer shows its source
Replies cite the documents they came from, so an agent or a customer can check them.
One knowledge base per business
Each tenant's documents, conversations and users are kept separate, with admin, agent and customer roles.
result
- Code on GitHub.
- The Ask console on this site uses the same idea at a smaller scale: answers grounded in documents, with the sources shown.