Ingestion pipeline
Connectors, parsing and chunking strategies tuned to your document types, kept in sync as content changes.
RAG Assistants & Chatbots
Answers grounded in your documents, with sources.
I build retrieval-augmented assistants that answer questions from your documents, wikis and knowledge bases: accurately, with citations, and with access controls that respect who is asking. Quality is measured with an evaluation suite, not guessed.
Sound familiar?
Problems this solves
Knowledge is scattered across SharePoint, PDFs, wikis and old tickets.
A generic chatbot hallucinates and can't show where an answer came from.
You're worried about sensitive documents reaching the wrong people.
What you get
Scope of work
A typical engagement covers the following, tailored to your stack and goals.
Hybrid search & reranking
Vector plus keyword search with reranking, so the right passages reach the model.
Assistant with citations
Grounded answers with links to sources, follow-up questions and guardrails against off-topic use.
Access control
Role-based filtering so each user only retrieves documents they are permitted to see.
Evaluation suite
Golden datasets with RAGAS and LLM-as-a-judge scoring to track answer quality release to release.
Delivery where people work
A web app, a Microsoft Teams bot with Adaptive Cards, or an API inside your product.
Use cases
Where this fits
- Internal assistant for policies, SOPs and HR questions
- Customer support assistant over help-centre content
- Sales and RFP assistant over past proposals
- Engineering assistant over technical documentation
Toolkit
Technology I use
- LlamaIndex
- LangChain
- Azure AI Search
- PgVector
- Qdrant
- Pinecone
- Rerankers
- RAGAS
- Teams Bot Framework
I pick tools to fit your environment and constraints, not the other way round. Models and vendors stay swappable.
Related experience
Built this before
From enterprise platforms I've built. Client names are withheld.
FAQ
Questions about rag assistants & chatbots
Will our documents be used to train a model?
No. Your documents are indexed in a store you control and sent to the model only as context for a question. Enterprise model endpoints (for example Azure OpenAI or AWS Bedrock) do not train on your data.
How do you measure whether answers are good?
We build a test set of real questions with expected answers and score every release on retrieval quality, faithfulness to sources and answer correctness, so improvements are measurable.
Talk to me about rag assistants & chatbots
Tell me what you're trying to build. In a free 30-minute call we'll work out whether AI is the right tool, what it would take, and whether I'm the right person to build it.