Document Summerization with GenAI for Compliance Team

Step 0: Sample Loan Agreement Document

File: sample-loan-agreement.pdf

Uploaded by: Compliance Officer Amit

Document Type: Term Loan Agreement

Sample Content (Excerpt):

1. Parties: ABC Bank (Lender) and XYZ Corp (Borrower)
2. Loan Amount: $20,000,000
3. Term: 5 years
4. Interest Rate: SOFR + 2%, compounded quarterly
5. Repayment: Quarterly installments
6. Collateral: Real estate properties at XYZ Corp HQ
7. Covenants:
   7.1 Borrower must maintain DSCR ≥ 1.25
   7.2 Borrower may request grace period up to 180 days
   7.3 Quarterly financial reports to be submitted within 30 days
8. Prepayment: Allowed with 2% penalty
9. Regulatory:
   AML compliance under FATCA & RBI KYC 2023 Master Circular
   Basel III exposure norms for large corporate borrowers
10. Action Items:
    Compliance verification of FATCA clause
    Risk team sensitivity analysis on revenue forecast

Step 1: Angular UI Upload

Actions:

• Compliance Officer Amit logs in (RBAC access).

• Uploads PDF via multipart form.

• Metadata submitted: document type, version, sensitivity.

Tables Updated:

1. Audit Table: audit_log

audit_id	event_name	document_id	performed_by	timestamp	details
A1	DOCUMENT_UPLOAD_INITIATED	null	Amit	2025-10-04 10:00	Upload initiated, metadata validated

Event Emitted: DOCUMENT_UPLOAD_INITIATED

Consumed By: AuditService

AI Guardrails: Input validation, file type/size, PII masking.

Step 2: Document Service - Save & Persist

Actions:

• File saved locally: /data/documents/sample-loan-agreement.pdf

• Metadata stored in document_metadata table.

Table: document_metadata

document_id	file_name	doc_type	uploaded_by	upload_date	version	sensitivity	storage_path	processed_flag	processed_date
UUID1	sample-loan-agreement.pdf	LoanAgreement	Amit	2025-10-04 10:01	1.0	Confidential	/data/documents/sample-loan-agreement.pdf	false	null

Audit Table Entry:

audit_id	event_name	document_id	performed_by	timestamp	details
A2	DOCUMENT_UPLOADED	UUID1	DocumentService	2025-10-04 10:01	File saved, metadata persisted

Event Emitted: DOCUMENT_UPLOADED

Consumed By: OCRService, AuditService

Step 3: OCR & Text Extraction

Actions:

• OCRService consumes DOCUMENT_UPLOADED.

• Extracts text and splits into chunks.

• PII masked or tokenized.

Table: document_text

chunk_id	document_id	text_chunk	pii_masked_flag
1	UUID1	"Parties: ABC Bank (Lender) and XYZ Corp..."	true
2	UUID1	"Loan Amount: $20,000,000; Term: 5 years..."	false
3	UUID1	"Covenants: 7.1 DSCR ≥1.25; 7.2 grace 180d..."	false

Audit Table Entry:

audit_id	event_name	document_id	performed_by	timestamp	details
A3	DOCUMENT_OCR_COMPLETED	UUID1	OCRService	2025-10-04 10:03	3 text chunks extracted, PII masked

Event Emitted: DOCUMENT_OCR_COMPLETED

Step 4: Indexing & Embeddings

Actions:

• IndexingService consumes DOCUMENT_OCR_COMPLETED.

• Generates embeddings for each chunk via OpenAI API.

• Stores embeddings in PGVector table.

• Redis caching for frequently accessed embeddings.

Table: document_embeddings (PGVector)

chunk_id	document_id	embedding_vector	created_at
1	UUID1	[0.12,0.56,...0.78]	2025-10-04 10:04
2	UUID1	[0.98,0.34,...0.11]	2025-10-04 10:04

Audit Table Entry:

audit_id	event_name	document_id	performed_by	timestamp	details
A4	DOCUMENT_INDEXED	UUID1	IndexingService	2025-10-04 10:05	3 chunks indexed, embeddings created

Event Emitted: DOCUMENT_INDEXED

Step 5: Summarization (Executive Summary)

Actions:

• SummarizationService consumes DOCUMENT_INDEXED.

• LLM generates structured summary with AI Guardrails (hallucination checks, citation, explainability, validation).

Table: document_summary

summary_id	document_id	executive_summary	key_obligations	risks	regulatory_mapping	clause_categorization	action_items	feedback_status
S1	UUID1	"Loan agreement: $20M, 5y, SOFR+2%, quarterly"	"DSCR ≥1.25; Quarterly reports; Prepayment"	"Grace 180d; Cross-border guarantees"	"AML: FATCA & RBI KYC; Basel III exposure"	"Covenants, Prepayment, Repayment"	"Verify FATCA; Risk revenue analysis"	Pending

Audit Table Entry:

audit_id	event_name	document_id	performed_by	timestamp	details
A5	DOCUMENT_SUMMARIZED	UUID1	SummarizationService	2025-10-04 10:06	Summary generated, AI guardrails applied

Event Emitted: DOCUMENT_SUMMARIZED

Step 6: Feedback Loop

Actions:

• Compliance officer reviews summary: Approve / Request changes.

• Updates document_summary.feedback_status.

Audit Table Entry:

audit_id	event_name	document_id	performed_by	timestamp	details
A6	FEEDBACK_UPDATED	UUID1	Amit	2025-10-04 10:07	Feedback: Approved

Event Emitted: FEEDBACK_UPDATED

Step 7: Approval & Notifications

Actions:

• ApprovalService consumes FEEDBACK_UPDATED.

• Updates document_metadata.processed_flag = true and processed_date.

• Notifies risk & compliance dashboards.

Audit Table Entry:

audit_id	event_name	document_id	performed_by	timestamp	details
A7	DOCUMENT_APPROVED	UUID1	ApprovalService	2025-10-04 10:08	Document approved, dashboards notified

Event Emitted: DOCUMENT_APPROVED

Step 8: Semantic Search & Retrieval

Actions:

• SearchService queries PGVector embeddings for semantic search.

• Returns top-k relevant chunks.

Audit Table Entry:

audit_id	event_name	document_id	performed_by	timestamp	details
A8	DOCUMENT_SEARCHED	UUID1	Amit	2025-10-04 10:09	Retrieved 3 similar chunks for query

Step 9: Event Flow Summary

Event Name	Produced By	Consumed By	Action Taken
DOCUMENT_UPLOAD_INITIATED	Angular UI/Controller	AuditService	Log user & metadata
DOCUMENT_UPLOADED	DocumentService	OCRService, AuditService	OCR extraction, log metadata
DOCUMENT_OCR_COMPLETED	OCRService	IndexingService, AuditService	Generate chunks, PII masking
DOCUMENT_INDEXED	IndexingService	SummarizationService, SearchService, AuditService	Embeddings created, cached
DOCUMENT_SUMMARIZED	SummarizationService	FeedbackService, AuditService	Executive summary generated, AI guardrails applied
FEEDBACK_UPDATED	FeedbackService	ApprovalService, AuditService	Feedback processed, retraining if needed
DOCUMENT_APPROVED	ApprovalService	AuditService	Final approval logged
DOCUMENT_SEARCHED	SearchService	AuditService	Semantic search logged

✅ Key Features Illustrated in This Walkthrough:

• Tables: document_metadata, document_text, document_embeddings (PGVector), document_summary, audit_log

• AI Guardrails: Input/output validation, hallucination checks, PII masking, citation, explainability

• Caching: Redis cache for embeddings

• Event-Driven Architecture: Full asynchronous flow with events emitted and consumed

• Audit: Each action is logged with timestamp, actor, document_id, event_name