RAG Assistant for a Foncier-Sector Client

Situation

A land-administration (foncier) company managed a large and growing corpus of property and parcel documents. Staff spent hours manually searching and cross-referencing these records to answer internal and customer questions.

Task

The client wanted a system that could answer natural-language questions about the corpus with traceable, citation-backed answers — not just a search box.

Action

I designed and shipped a retrieval-augmented generation (RAG) agent over the document corpus:

• Chunking & indexing: documents parsed, split into semantically coherent chunks, embedded, and stored in Qdrant as the vector store.
• Retrieval: top-k vector search with metadata filtering to scope results to the right document types and dates.
• Agent: a LangChain/LangGraph agent that plans between retrieval, reading source documents, and composing answers, keeping the model on-task with a small toolset.
• Answering: final answers with inline references back to the source chunks, so every claim could be verified.

Result

• Cut the time to answer a typical multi-document question from 15–20 minutes to under a minute.
• Citations meant staff could verify answers instead of re-searching manually — essential in this regulated industry.
• The architecture became a reusable blueprint the client could repurpose for other corpora.

Confidentiality note

The underlying technical and data specifics of this regulated project stay intentionally general here. Client name omitted at request of engagement — the same posture used across my other case studies.

Stack highlights

• Qdrant as the vector database (hosted)
• Voyage AI for embeddings, Anthropic Claude for generation
• LangGraph for the agentic control flow
• Docker for reproducible deployment for the client's infrastructure

The exact same RAG pattern powers the chat widget on this very site.