RAG Chatbot Development for Public Documents

"Retrieval-Augmented Generation (RAG)" chatbot: official public documents are ingested, chunked, embedded, and stored in a vector database. When a user asks a question, the system retrieves the most relevant passages and uses an LLM to generate an answer grounded in that retrieved content — with every factual claim cited to a specific source and date. If the system can't find a reliable source, it says so instead of guessing.

## Key Features
- "Grounded, cited answers" — every claim traceable to a specific document, section, and date
- "Three answer types" — quick factual lookups, analytical/statistical summaries (trends, comparisons), and step-by-step procedural guidance
- "Hallucination guardrails" — a verification step checks generated claims against retrieved source material before an answer is shown
- "Graceful "no answer" handling" — out-of-scope or low-confidence questions get a plain "no sourced answer available" instead of a confident guess
- "Multi-channel access" — a full-featured web chat interface plus at least one lightweight messaging channel (e.g. WhatsApp) for lower-bandwidth users
- "Source-conflict handling" — outdated or superseded information is clearly flagged, with current versions favored by default
## Scope of Work (MVP)
- Data ingestion pipeline for one initial public-information domain (regulatory/procedural documents + statistical releases)
- Chunking and metadata strategy that treats prose text and structured/statistical data differently
- Hybrid retrieval (keyword + semantic search) with metadata filtering by document type, date, and status
- LLM-based answer generation with inline citations and a grounding/verification check
- Web chat interface with expandable citations and table support for comparisons
- One messaging-channel integration with citation display adapted to that channel
- Basic audit logging of questions and answers for quality review (no personal data required to use the assistant)

## Suggested Tech Stack
- "Vector database:" pgvector, Weaviate, or Qdrant
- "Embedding model:" a current-generation multilingual embedding model
- "LLM:" opensource model with strong grounded-generation performance
- "Orchestration:" a RAG framework (e.g. LlamaIndex) or a custom retrieval router
- "Structured data layer:" PostgreSQL for statistics/tabular data, queried directly rather than only via embeddings
- "Hosting:" cloud VM/container hosting with a managed database option

*(Stack is indicative — open to the developer's recommendation based on budget and infrastructure constraints.)*