01 — Project
Procurement AI
Intelligent RFP Evaluation Platform
AI vendor proposal evaluation that scores compliance, extracts commercial terms, and ranks winners.
1024
Embedding Dim
1000
Chunk Size (Chars)
0.3
Search Threshold
GPT-4o
AI Model
20K
Vector Scroll
0
Temperature
02 — Evaluation Pipeline
1. Repository
Create repository & Qdrant collection
2. Ingestion
Upload RFP & quotes (PDF, DOCX, XLSX)
3. Extraction
Parse text & tables · Hybrid header detect
4. Chunking
Smart chunking (1000 chars, 100 overlap)
5. Embedding
BGE-large (1024d) · Upsert to Qdrant
6. Retrieval
Semantic search (requirement chunks)
7. Compliance Audit
GPT-4o + Regex rules · Safety validations
8. Financial Analysis
Extract & normalize commercial terms
9. Ranking
Score, rank & tie-break (disqualified last)
10. Winner Report
Generate matrices, remarks & winner
03 — Business Impact
The Problem
- Manual evaluation is slow & inconsistent
- Unstructured docs with inconsistent layouts
- Safety-critical spec violations
- Financial line items are hard to normalize
- Hard to justify winner to stakeholders
The Solution
- Ingest & parse all documents
- Hybrid RAG + vector search
- Deterministic compliance audit with hard rules
- Financial extraction & normalization
- Structured matrices & winner report
The Outcome
- Accurate compliance scoring
- Financial transparency across vendors
- Defensible ranking & audit trail
- Hours of manual work replaced per RFP
- Deterministic winner recommendation
04 — Key Metrics & Parameters
Chunk Size
1,000
Characters
Embedding Dim
1,024
Dimensions
Search Threshold
0.3
Cosine Similarity
Scroll Capacity
10 K – 20 K
Vector Points
Compliance Scoring
1 / 0.5 / 0 / −100
Full / Partial / Non / Critical
Temperature
0
Deterministic
05 — System Architecture
1. Repository
Create repository & Qdrant collection
2. Ingestion
Upload RFP & quotes (PDF, DOCX, XLSX)
3. Extraction
Parse text & tables · Hybrid header detect
4. Chunking
Smart chunking (1000 chars, 100 overlap)
5. Embedding
BGE-large (1024d) · Upsert to Qdrant
6. Retrieval
Semantic search (requirement chunks)
7. Compliance Audit
GPT-4o + Regex rules · Safety validations
8. Financial Analysis
Extract & normalize commercial terms
9. Ranking
Score, rank & tie-break (disqualified last)
10. Winner Report
Generate matrices, remarks & winner
Core AI Components
GPT-4o
(Deterministic)
Qdrant
(Vector Search)
AWS S3
(Document Store)
1. Repository
Create repository & Qdrant collection
2. Ingestion
Upload RFP & quotes (PDF, DOCX, XLSX)
3. Extraction
Parse text & tables · Hybrid header detect
4. Chunking
Smart chunking (1000 chars, 100 overlap)
5. Embedding
BGE-large (1024d) · Upsert to Qdrant
6. Retrieval
Semantic search (requirement chunks)
7. Compliance Audit
GPT-4o + Regex rules · Safety validations
8. Financial Analysis
Extract & normalize commercial terms
9. Ranking
Score, rank & tie-break (disqualified last)
10. Winner Report
Generate matrices, remarks & winner
Core AI Components
GPT-4o
(Deterministic)
Qdrant
(Vector Search)
AWS S3
(Document Store)
06 — Core Features
Technical Compliance Matrix
Requirement-level scoring
Financial Breakdown Matrix
Normalized commercial terms
Vendor Ranking & Winner
Deterministic tie-break
Safety Rule Validation
LLM + regex overlays
Commercial Terms Extraction
Price, delivery, line items
Disqualification Detection
Critical fail → last place
Repository Isolation
(per RFP)
Audit Trail & Remarks
Per-vendor justification
Export Reports
(HTML / Excel)
Multi-format Documents
PDF, DOCX, XLSX
Currency Parsing
Multi-currency normalize
Multi-section Specifications
Nested requirement handling
07 — Engineering Innovations
Hybrid Header Detection
Detects correct table headers even in multi-row or messy spreadsheets.
Atomic Table Chunking
Table rows kept atomic to preserve pricing & item relationships.
Dual Validation (LLM + Regex)
GPT reasoning combined with hard regex rules for numeric safety specs.
Financial Normalization
Maps vendor terminology to standard line items for fair comparison.
Deterministic Scoring
Compliance rules ensure reproducible and auditable results.
Repository Isolation
Per-RFP collections for secure and accurate evaluations.
08 — Code Highlight
async def perform_comparison(repo_id: int):
- 1Load repository & vendor documents
- 2Retrieve requirement chunks from Qdrant
- 3For each vendor quote:
- 4Run GPT-4o compliance audit
- 5Apply regex safety validations
- 6Extract & normalize financial terms
- 7Build technical & financial matrices
- 8Calculate compliance score
- 9Rank vendors (disqualified last, score, price, name)
- 10Generate HTML report & return winner
09 — Handled Edge Cases
Multi-row Headers
Critical Numeric Safety Specs
Malformed Documents
Currency Parsing
Missing Totals
Inconsistent Terminology
LLM JSON Recovery
Large Document Handling
Multi-section Specifications
Interested in how we built it?
AI vendor proposal evaluation that scores compliance, extracts commercial terms, and ranks winners.