
Project Overview#
I engineered a Retrieval-Augmented Generation (RAG) system designed to bridge the gap between static corporate documentation and interactive AI. The system autonomously monitors internal directories, processes unstructured data, and provides a conversational interface for complex policy queries.
1. Automated ETL Pipeline (Data Sync)#
Instead of manual data entry, I built an automated ingestion layer that keeps the AI’s “brain” current:
- Source Integration: Monitors Google Drive repositories for new or updated policy documentation.
- Vectorization: Utilizes Google Gemini Embeddings to convert text into high-dimensional vectors.
- Database Management: Automates the flushing and re-indexing of a Pinecone Vector Database to ensure 100% data consistency.
2. Intelligent Document Processing#
To ensure the AI handles large documents accurately, the engine employs sophisticated text-handling logic:
- Recursive Chunking: Breaks down large PDF/Docx files into optimized segments with a 100-character overlap to maintain semantic context across chunks.
- Metadata Tagging: Ensures every piece of retrieved information is traceable back to its source.
3. Agentic Logic & Retrieval#
The front-end interface utilizes an “Agentic” model rather than a simple prompt:
- Contextual Memory: Implemented a Window Buffer Memory system, allowing the AI to remember the last several turns of a conversation for fluid follow-up questions.
- Semantic Search Tooling: The agent is equipped with a custom-tooled vector retriever, allowing it to perform “lookups” in real-time before formulating a response.
- Zero-Shot Guardrails: Engineered strict system prompts to prevent the AI from speculating; if the answer isn’t in the verified docs, it will not provide a false answer.
Technical Architecture Summary#
| Component | Technology | Role |
|---|---|---|
| Model | Gemini 2.0 Flash | Primary Reasoning Engine |
| Vector Store | Pinecone | Semantic Long-term Memory |
| Data Source | Google Drive API | Source of Truth |
| Orchestration | Event-Driven Middleware | Pipeline Logic & API Glue |
| Embeddings | Google PaLM/Gemini | Text-to-Vector Transformation |
