Project Overview
A small US consulting firm had an organizational risk most growing professional services businesses will recognize: nearly all of the firm’s institutional knowledge lived inside one senior consultant. When she was occupied or unavailable, junior staff struggled to answer routine client questions with confidence, and the senior consultant spent a disproportionate share of her time fielding internal interruptions rather than doing billable work. We built a RAG system trained on five years of her documented processes, past proposals, meeting notes, and client communication history that gave the whole team reliable access to that knowledge on demand.
The Challenge
The knowledge concentration problem was practical and immediate. Junior consultants would either guess at answers to client questions, producing inconsistent or incorrect responses, or interrupt the senior consultant constantly, creating a bottleneck that limited the firm’s capacity to take on more work. There was no documentation system, no knowledge base, and no structured way to transfer what the senior consultant knew to the rest of the team.
- Five years of processes, proposals, and client correspondence existed only in unstructured form across Word documents, PDFs, and email threads
- Junior staff had no reliable way to find the right answer to client questions without escalating to the senior consultant
- The senior consultant was spending significant hours per week on internal interruptions that pulled her away from client-facing work
- The business was effectively bottlenecked around one person’s availability
Our Solution
We built a RAG system that indexed five years of the senior consultant’s documented knowledge into a queryable knowledge base that junior staff could access in plain English. Instead of searching through folders or interrupting a colleague, a junior consultant could type a question and receive a specific, context-aware answer drawn from the firm’s own documents and history.
- Document Extraction and Cleaning: We extracted and preprocessed content from Word documents, PDFs, and exported email threads, cleaning inconsistent formatting and resolving duplicate content before any indexing began.
- Chunking and Embedding: Documents were chunked by logical section and embedded using a text embedding model, then stored in a vector database with metadata tagging by document type, client name, and date range.
- RAG Query Interface: A simple plain-English interface allowed junior staff to ask questions like “How do we handle scope changes for retainer clients?” or “What was our approach on the Henderson project?” and receive answers grounded in actual firm documents, with source citations attached.
- Confidence Flagging: Responses below a retrieval confidence threshold were flagged visually so staff knew when to seek additional verification rather than treating an answer as definitive.
Technical Approach
The retrieval layer used semantic vector search over chunked document embeddings, with a reranking step to improve precision on queries that matched multiple document segments. The generation layer was prompted to answer only from retrieved context and to cite the source document for every factual claim, keeping outputs grounded and verifiable. The system was hosted in the firm’s existing cloud environment with access controls limiting document visibility to staff with appropriate permissions.
Results & Impact
- Five years of previously unstructured documents indexed and accessible to the full team for the first time
- Junior staff became self-sufficient on routine client queries, resolving questions that had previously required senior consultant involvement
- Hours saved per week for the senior consultant, recovered from internal interruptions and reallocated to billable client work
- Institutional knowledge preserved in a durable, searchable format that reduced the firm’s operational risk from any single person’s absence
- Onboarding time for new junior consultants reduced, as the knowledge base gave them a reliable reference from day one
Lessons Learned
The document extraction and cleaning phase took longer than the modeling work. Five years of Word documents, PDFs, and email exports arrived in inconsistent formats with varying quality. Investing in that foundation was essential: a RAG system is only as useful as the quality and completeness of the documents it retrieves from, and rushing that step would have produced a system that gave confidently wrong answers, which would have been worse than no system at all.
The confidence flagging feature also turned out to be important for trust. Junior staff needed to know when the system was uncertain so they could decide whether to escalate rather than relying on a low-confidence answer. That transparency made the tool more trustworthy, not less, and drove higher adoption than a system that always produced an answer regardless of retrieval quality would have.