Compliance Report Generator
A full-stack compliance agent that processes regulatory documents (GDPR, internal policies) and makes them queryable through multiple retrieval strategies. Users upload PDFs, ask questions, and receive answers grounded in the document — with every response gated through PII detection, prompt safety checks, and human approval before being saved to memory.
Pipeline
- PDF uploaded and parsed into chunks using pdfplumber (text + tables)
- Chunks embedded and stored in Qdrant for semantic vector search
- Entities and relationships extracted into Neo4j knowledge graph
- User query routed to the selected agent mode
- Safety checks run (PII detection + prompt guardrails) before any LLM call
- Answer shown for human review — approved answers saved to Qdrant memory
Key Components
Multi-Modal Retrieval — vector similarity (Qdrant) finds semantically relevant chunks; Neo4j graph queries surface entity relationships that keyword search misses.
Three Agent Modes
- Memory-Aware QA: retrieves document context with Neo4j graph memory injected into the prompt, role-specific instructions for legal analyst, policy researcher, or compliance officer
- Tool Agent: LangChain ZERO_SHOT_REACT agent with 5 tools — risk scoring, compliance lookup, live news, summarization, and compliance score
- Context-Aware Chain: role-based QA chain with custom PromptTemplates per user type
Security Layer
- PII detection (emails, phone numbers, SSNs) before any LLM call
- Keyword-based prompt guardrails blocking harmful or unethical queries
- Ollama-based self-hosted safety classifier for air-gapped environments where prompts cannot leave the machine
- Consistent enforcement across CLI, agents, and Streamlit UI
Human-in-the-Loop (HITL)
- Streamlit UI: session state pauses execution, shows answer and source chunks, waits for approve or regenerate
- CLI: programmatic HITL wrapper for batch workflows — runs chain, prints answer with sources, prompts for approval before saving to memory
LangGraph Ingestion Workflow — checkpointable pipeline (ingest, embed, graph, memory) that resumes from any failed step without reprocessing earlier stages.
What I Learned
- How to combine vector search and knowledge graphs for richer document retrieval — semantic similarity finds relevant chunks, graph queries find entity relationships that keyword search misses
- LangGraph for stateful, fault-tolerant workflows — nodes can fail and resume without reprocessing earlier steps
- Building practical guardrails: PII regex detection and prompt classification before any LLM call, with a self-hosted Ollama fallback for privacy-sensitive deployments
- Human-in-the-loop patterns in Streamlit using session state to pause execution and wait for user approval before writing to memory
- Multi-agent orchestration: separating memory-aware QA from tool-calling agents and designing a tool registry that makes agent capabilities inspectable and extensible