Planit Consulting

Before anything could be searched, the archive had to be sorted. A Python pipeline classifies each .docx into one of 14 types across 3 tiers, from primary documents down to metadata-only files, so retrieval later knows what it is pulling.

• Classifies 18,429 .docx into 14 document types across 3 tiers (primary, secondary, metadata-only).
• The 514 documents that could not be reliably typed are split off and handled separately, instead of polluting the searchable set.
• Reliable typing is the foundation: it lets each document type get the chunking and retrieval treatment it actually needs.

Each classified document is split into passages with section detection, then turned into vectors. We used OpenAI text-embedding-3-large for precision, with the spend managed at the chunk level so the build stayed economical at scale.

• Semantic chunking with section detection produces 95,592 chunks from 15,801 documents.
• Embedding via OpenAI text-embedding-3-large (3072 dimensions), chosen for precision over cheaper alternatives.
• Oversized chunks get separate handling, and cost is controlled at the chunk level rather than re-embedding everything blindly.

The embeddings live in a persistent ChromaDB vector store, served by a FastAPI backend. This is the layer that turns a question into the handful of passages most relevant to it, in a fraction of a second.

• ChromaDB vector store with persistent storage holds all 95,592 chunks and their metadata.
• FastAPI backend exposes REST endpoints with CORS for the chat UI to call.
• Semantic retrieval finds passages by meaning, so a query surfaces the right precedent even when it does not share the exact wording.

A React TypeScript chat interface, named Planit AI, is where consultants actually work. It is built for the way they ask questions, one follow-up after another, with a sidebar that keeps every line of enquiry organised.

• React TypeScript UI with a sidebar that groups conversations, plus rename and delete.
• Conversations and messages persist in SQLite, so a consultant can pick up an earlier line of questioning.
• Designed around iterative questioning, the way consultants narrow in on the precedent they need.

Claude reads the retrieved passages and writes the answer, with citations back to the original .docx. The citation is the point: every answer links to its source so a consultant can open the document and verify it.

• Claude answers from the retrieved passages, grounded in the firm's own documents rather than general knowledge.
• Every answer carries source citations back to the original .docx for verification.
• Consultants get the precedent in seconds and can always check it against the file it came from.