cosmic-architecture

byparesh ms

Role: You are an expert AI Systems Architect specializing in Intelligent Document Processing (IDP) and Knowledge Graphs. Your goal is to design the technical architecture for EvidenceGraph AI, a system that automates government tender evaluation with 100% auditability. Core Task: Develop a technical blueprint and initial code structure that implements a "Trust-First" evaluation engine. 1. Data Schema & Knowledge Graph Design Define a JSON schema for the Evidence Graph. It must link four nodes: Tender_Criterion, Extracted_Data_Point, Source_Metadata (Document ID, Page #, Bounding Box coordinates), and Evaluation_Status. Explain how to maintain a "Chain of Custody" for every extracted value so a human can click a decision and see the exact highlighted text in the source PDF. 2. Document Processing Pipeline Outline a Python-based workflow using tools like Unstructured.io, PyMuPDF, or LayoutLMv3 to handle hybrid documents (scanned images + digital text). Create a logic for Value Normalization. (e.g., converting "5 Crores," "50,000,000," and "500 Lakhs" into a standard integer for rule-based comparison). 3. The Evaluation Engine Logic Write a Python function evaluate_bidder(criteria_list, extracted_data) that uses: Deterministic Rules: For hard thresholds (Financial turnover, years of experience). Probabilistic Matching: Using semantic similarity for complex criteria (e.g., "Has the bidder done 'similar' infrastructure work?"). Define the "Confidence Threshold" logic: If OCR confidence is <85% or semantic matching is <0.7, the status must be set to NEEDS_MANUAL_REVIEW. 4. Explainability & Audit API Design an API endpoint /get-decision-rationale/{bidder_id} that returns a natural language explanation alongside the evidence graph data. Example Output: "Rejected: Technical Criterion 4.2. Extracted: 'ISO 9001'. Required: 'ISO 27001'. Source: Certifications.pdf, Page 12." 5. Frontend "Human-in-the-Loop" UI Concept Describe the UI requirements for an Evaluator Dashboard. How should the "Split-Screen" view work (Decision Logic on the left, PDF Viewer with highlights on the right)? Output Format: Provide the response with high-level architectural diagrams (text-based), core Python class structures, and a sample JSON output for a single "Evidence Graph" node. Implementation Tips for Your Project For the Graph: Consider using a Vector Database (like Pinecone or Weaviate) combined with a Relational Database (PostgreSQL). The Vector DB handles the "similar project" searches, while Postgres stores the hard facts and audit trails. For OCR: If you are dealing with government-style scanned documents, look into Google Cloud Document AI or Azure Form Recognizer; they are much more resilient with tables and stamps than standard Tesseract. The "Evidence" part: Ensure your backend saves the Coordinates (x, y, w, h) of the text found on the page. This is what makes the system "traceable" for the user. Next Steps for Development Phase 1: Build a parser that turns a sample Tender PDF into a JSON list of requirements. Phase 2: Build the "Evidence Linker" that finds a requirement (e.g., "Turnover") in a Bidder's Balance Sheet. Phase 3: Create the UI that shows the "Red/Green" status with a link to the page number.