Build a RAG chatbot using your documents and website

Turn your documents and website into a reliable AI assistant that answers with evidence—not a generic chatbot that guesses.

I will build and evaluate one Retrieval-Augmented Generation system tailored to your business knowledge, users and questions.

TECHNICAL APPROACH

✓ Define the use case, answer policy and success criteria
✓ Audit document quality, structure, duplicates and metadata
✓ Parse PDF, DOCX, TXT or selected website pages
✓ Apply structure-aware or semantic chunking with justified overlap
✓ Create dense embeddings and metadata filters
✓ Add sparse BM25 retrieval when hybrid search is useful
✓ Store vectors with Qdrant, ChromaDB or FAISS
✓ Combine dense and sparse results using RRF when appropriate
✓ Rerank passages with a cross-encoder or ColBERT
✓ Configure the LLM to answer from retrieved evidence
✓ Add source citations and an “insufficient evidence” response
✓ Apply prompt-injection and off-topic guardrails
✓ Test retrieval, faithfulness, relevance and latency
✓ Log failed questions and knowledge gaps

EVALUATION

I will create a test set from your representative questions. Depending on the project, evaluation may include Recall@K, MRR, context precision/recall, answer relevance and faithfulness using RAGAS or equivalent checks.

The system is designed to reduce unsupported answers, but no responsible RAG solution should promise zero hallucinations.

YOU WILL RECEIVE

• One custom RAG chatbot for one use case
• Processed and indexed knowledge base
• Hybrid retrieval and reranking when justified
• One LLM and vector-database integration
• Answers with source citations
• Basic Streamlit demonstration interface
• Documented Python source code
• Evaluation results and test questions
• README and usage instructions
• One revision and seven days of support

BASE OFFER SCOPE

• Up to 20 documents or 200 pages
• PDF, DOCX, TXT or selected website pages
• One knowledge base and primary language
• One LLM provider and vector database
• Up to 20 evaluation questions

For scanned-document OCR, SQL or API connections, website embedding, authentication, production deployment or larger knowledge bases, please message me for a tailored scope.

Please provide your documents or URLs, intended users, 10–20 representative questions, preferred answers, language, tone, restricted topics and provider preference.