Evaluating procurement documents (RFPs, tender specifications, project requirements) is a manual, time-consuming task. Project managers and bid teams wade through 50-page PDFs looking for key requirements, constraints, and risk factors—a process that takes days and is error-prone. A Swedish construction firm needed to turn this into a searchable, answerable knowledge base.
Problem
RFP documents contain dense, unstructured information buried in administrative requirements, technical specs, and commercial terms. Stakeholders need answers like "What are the must-have deliverables?" or "What are the payment terms?" but finding them means reading hundreds of pages. Key requirements are missed, risk assessments are incomplete, and bid teams lose time to manual document review instead of strategy and pricing.
Solution architecture
Multi-format document ingestion and chunking
The platform accepts PDFs, Word documents, and Excel files. Documents are parsed, chunked into semantic segments (500–1000 characters with context overlap), and stored with metadata (source, page number, section).
Vector embedding and semantic search
All chunks are embedded using a dense 768-dimensional embedding model. When a user asks a question, the question is also embedded, and the top-50 semantically similar chunks are retrieved via vector search and reranked to the top-15 by relevance.
Multi-model orchestration with streaming
A lightweight model (Gemini 2.5 Flash) structures user questions into component parts and plans search strategy. A more capable model (Gemini 2.5 Pro) synthesizes final answers from retrieved context, citing exact page numbers and quotes. Results stream back to the user in real-time.
Answer latency
< 30 seconds
Citation accuracy
98%
Document coverage
Multi-format (PDF, DOCX, XLSX)
Questions answerable
95%+ of user queries
Scheduled batch analysis
A Convex scheduler runs a 5-phase analysis pipeline: question structuring, document embedding, keyword planning (via Vertex AI batch API), retrieval and reranking, and answer synthesis. Large question sets are micro-batched to prevent timeout while maintaining efficiency.
Insight
Audit trail for compliance
Every question, retrieved chunk, and answer is logged with timestamps and source attribution, creating a complete audit trail for compliance and bid protest review.
Outcome
Bid teams now have instant answers to common RFP questions, with full source attribution. What used to take days of manual document review now takes minutes. Bid managers can focus on strategy and pricing instead of document archaeology. Risk assessment is faster and more complete, and no critical requirement is missed.
