RAG4AO
Augmented search for tenders — intelligently compare CCTP specification documents by theme
Discuss a similar projectRéponse courte
RAG4AO is a retrieval-augmented (RAG) search platform developed by Digit-AI for teams responding to public tenders. It automatically analyzes CCTP specifications across 12 thematic dimensions (social clauses, regulations, services…) using 12 distinct FAISS vector indexes. The result: the time to find similar tender precedents drops from 2–4 hours to under 2 minutes (−98%), with analysis coverage multiplied by 4 to 6.
Context & Challenges
A manual, time-consuming process dependent on individual memory
Teams responding to public tenders handle hundreds of heterogeneous CCTP specifications (PDF, DOCX, TXT). Each new tender must be manually compared with previous ones to identify similarities, build on past responses and assess whether a bid is worthwhile — an entirely manual, time-consuming and untracked process.
Inefficient document search
No tool makes it possible to quickly retrieve past CCTP specifications similar to a new tender.
Knowledge loss
Expertise is held by a few specialists, with no structured capitalization or traceability.
Superficial analysis
Comparison is done on the tender title or subject, with no thematic granularity (social clauses, regulations, services…).
Longer decision time
Between a tender being published and the go/no-go decision, teams waste valuable time searching for previous tenders.
Functional scope
Tender response managers, project leads, purchasing/sales management. Covered processes: batch indexing of the corpus, similarity search, history consultation, multi-tenant administration.
Key constraints
Partitioned multi-tenant architecture · Strict GDPR (no external download) · Volume of hundreds to thousands of CCTP specifications · Lightweight Docker infrastructure compatible with Railway and on-premise.
Objectives & Scope
Three levels of objectives, each tied to a measured KPI
Strategic objective
Reduce tender decision time by providing fast, structured document comparison organized by theme.
Operational objective
Automate the thematic analysis of CCTP specifications across 12 business dimensions: title, summary, buyer, social/environmental clauses, scope, stakes, sites, services, regulations, criteria, resources, duration.
Technical objective
Deploy a containerized multi-tenant platform (Docker) with FAISS vector search, with no service disruption and without downloading the original documents.
Tech Stack
A multi-thematic RAG architecture with 12 distinct FAISS indexes
Frontend
Backend
LLM & Embeddings
Vector search & NLP
Documents & Data
Infrastructure
Differentiating pattern — multi-thematic RAG: unlike a classic single-index RAG, RAG4AO maintains 12 distinct FAISS indexes (one per theme). Each document is broken down into 12 facets by the LLM, each facet is vectorized and indexed separately. Search is performed theme by theme, offering a granularity and explainability impossible with a monolithic index.
Architecture Diagrams
Application-level overview and detailed indexing pipeline
Results, Gains & ROI
Measured gains across the entire tender response cycle
| Indicator | Before RAG4AO | After RAG4AO | Gain |
|---|---|---|---|
| Time to find tender precedents | 2–4 hours per tender (manual) | 30 seconds to 2 minutes | ~98% reduction |
| Thematic coverage | 2–3 criteria (title, subject, amount) | 12 themes analyzed automatically | ×4 to ×6 |
| Search response time | N/A | 10 to 30 seconds | Ultra fast |
| Usable documents | Dozens (individual memory) | Hundreds to thousands (full corpus) | ×10 to ×100 |
| Batch indexing capacity | N/A | ~500–700 documents/hour | New process |
Qualitative gains
Explainability
Results structured by theme: users understand why two documents are similar (same social clauses, same services, same regulations…).
Knowledge capitalization
The indexed corpus builds a lasting organizational memory, independent of individuals and accessible to the entire team.
Error reduction
Automatic analysis across 12 dimensions prevents oversights: undetected environmental clauses, missed regulatory requirements…
Data sovereignty
Ollama option (local LLM such as Mistral or Deepseek) for 100% on-premise processing, with no data leaking to third-party APIs. GDPR-compliant.
Scalability
Multi-tenant architecture enabling onboarding of new entities without redeployment. SuperAdmin / CompanyAdmin / User roles.
Easy adoption
Modern React 19 interface, dark/light mode, Tailwind CSS 4. An experience close to mainstream search engines.