Years of proposals, safely searchable — half the research time.
Professional services firm with a heterogeneous corpus of past proposals — PDF, Word, PowerPoint, Excel — mixed public + client-sensitive content.
The proposal teams re-use years of past work — solutions, pricing references, case studies — every time a new client opportunity comes in. The knowledge exists, scattered across thousands of documents. Finding it, safely, was the bottleneck.
THE PROBLEM
Two constraints made this harder than a typical enterprise search project:
- Multi-format corpus. PDFs, Word docs, PowerPoint, Excel — years of content in wildly different structures. Native search across those formats was slow and shallow.
- Sensitive content mixed in. Many past proposals contained classified or client-confidential details that absolutely could not be surfaced in new documents. Manual review to check for leaks was slow AND risky.
The team needed a system that behaved like a senior consultant who had read every proposal — but knew exactly what could and could not be shared.
OUR APPROACH
We built a classification-aware RAG assistant. Each document chunk is tagged with sensitivity level during ingestion. Vector search filters retrieval by the user's clearance level. A separate LLM-based post-processing pass detects and masks any leaked sensitive content before the response reaches the user.
Users ask contextual questions; the system returns synthesised answers grounded in past proposals, with source citations for verification. If they don't have clearance for a source, that source never enters their retrieval results.
KEY TECHNICAL DECISIONS
- Classification-aware retrieval. Sensitivity tag per chunk; vector search filters by user clearance before generation.
- Hybrid semantic + BM25 search. Proposal terminology has specific jargon; BM25 catches exact-match terms while semantic covers meaning.
- Structure-preserving chunking. Sections, tables, and pricing blocks kept together during chunking — improves relevance and citation quality.
- Post-processing redaction pipeline. LLM-based detection layer catches sensitive content leaks in generated responses before display — belt-and-braces on top of the retrieval-level filtering.
- Full provenance tracking. Every answer links to source documents so users can verify context.
RESULTS
- Proposal research time cut ~50% on typical opportunities.
- Zero known sensitive-content leaks post-deployment. The classification-aware filtering + redaction combined layers are held up under audit.
- Proposal teams became more consistent across geographies — the same institutional knowledge is now reachable regardless of who's writing.
Got a document corpus with mixed sensitivity levels?
— A few seconds to fill in a form and we'll do the rest!
Discovery call, 45 minutes. Bring your document formats, your classification scheme, and the questions your team wastes time answering. We'll tell you where RAG helps, where it doesn't, and what a compliant pilot looks like.