AI RAG

01Project Overview

02The Challenge

The answers existed — in a shared drive, a wiki, a folder of PDFs, and several years of closed tickets. Finding them was the problem. The gaps were clear:

• Keyword search returned a list of files, not an answer, leaving people to read through them.
• Staff asked colleagues instead of searching, so a handful of experts became bottlenecks.
• A model on its own had no access to internal knowledge and confidently filled in the blanks.
• Superseded documents sat beside current ones with nothing to distinguish them.
• Some material was restricted, so a single open search index was never an option.
• Nobody could verify where an answer came from, which killed trust in the results.

The opportunity was to make the existing knowledge answerable — and to make every answer checkable.

03Our Approach

Retrieval quality decides whether a RAG system is useful, so that's where the work went:

• Inventoried every source and prioritized the documents people actually ask about.
• Built an ingestion pipeline that parses and chunks with structure intact — headings, tables, and clauses stay whole.
• Combined semantic and keyword search, then reranked, so both meaning and exact terms are covered.
• Required the model to answer only from retrieved passages and cite each one.
• Enforced permissions at retrieval time, so results are filtered to what each user may see.
• Re-index on change and prefer the current version when documents supersede each other.
• Scored retrieval against a real question set and tuned chunking and ranking against the results.

04Key Features

05How It Works

Documents are prepared once and kept fresh. At question time the system retrieves candidate passages, reranks them, and lets the model answer only from what it found.

INGESTION  (runs on change)                QUERY TIME
 ┌──────────────────────────────┐        ┌──────────────────────────────┐
 │ PDFs · Docs · Wiki · Tickets │        │ User question                │
 │            ▼                 │        │            ▼                 │
 │ Parse   keep structure       │        │ Retrieve vector + keyword   │
 │ Chunk   headings, tables     │        │ Filter   user permissions   │
 │ Embed   store vectors        │───────▶│ Rerank   best passages first │
 │ Index   vector + keyword     │        │ Answer   with citations     │
 └──────────────────────────────┘        └──────────────┬───────────────┘
                                                        │
                                    passage found       │      nothing relevant
                                                        ▼
                                          "Not covered in the documents"
                                          — logged as a knowledge gap

06Problems Solved

07Grounded by Design

08Results & Impact

09Conclusion

10Project Summary

11Technologies Used