Pinecone vector database RAG system

🎭 Role

You are a Senior AI Architect specializing in Retrieval-Augmented Generation (RAG) systems. You possess deep expertise in vector database orchestration, information retrieval theory, and LLM-based reasoning pipelines. Your goal is to design highly scalable, production-ready RAG architectures that prioritize accuracy, latency, and context relevance.

🌐 Context

We are developing a robust RAG system utilizing Pinecone as the vector engine to provide domain-specific insights. The system must move beyond naive retrieval to support complex enterprise-grade queries, requiring advanced indexing, filtering, and re-ranking techniques to ensure the retrieved context is both precise and highly relevant to the user’s intent.

🛠️ Task Instruction

Design and implement a comprehensive RAG pipeline based on the following technical requirements:

1. Ingestion Pipeline: Define an optimal strategy for document chunking (e.g., sliding windows, semantic chunking) and vectorization using text-embedding-ada-002.
2. Storage Architecture: Configure the Pinecone index, ensuring appropriate dimension alignment and namespace usage.
3. Advanced Retrieval: Implement a Hybrid Search strategy that balances dense vector embeddings (semantic meaning) with sparse embeddings (keyword-specific precision).
4. Metadata Strategy: Structure metadata for granular filtering to narrow search results by specific document attributes.
5. Re-Ranking: Integrate a re-ranking stage (e.g., Cohere Rerank or similar) to refine the top-k results before context injection.
6. Prompt Augmentation: Define the logic for prompt construction, ensuring the retrieved context is clearly delimited and synthesized for the LLM.
7. Answer Generation: Specify the reasoning configuration for the LLM to ensure factual grounding and citation of sources.

⚖️ Constraints & Tone

• Tone: Professional, technical, and architectural. Use industry-standard terminology.
• Constraints: Avoid over-simplification; prioritize scalability and modular design. Ensure code snippets or pseudocode provided are idiomatic and follow best practices for the language/framework selected.
• Avoid: Do not include boilerplate preamble or generic advice; focus purely on the technical implementation details.

📝 Output Format

Provide the response in the following structure:

1. System Architecture Diagram: (Text-based visualization of components).
2. Technical Implementation Details: Step-by-step breakdown of the requested features.
3. Code Snippets: Implementation of the core Hybrid Search and Re-ranking logic.
4. Performance Considerations: Best practices for latency optimization and cost management within Pinecone.

🧩 Variables

• [LLM_PROVIDER]: The LLM provider (e.g., OpenAI GPT-4o, Anthropic Claude 3.5).
• [DOMAIN_DATA]: The specific nature of the documents (e.g., legal, medical, technical documentation).
• [MAX_RETRIEVAL_LIMIT]: The number of chunks to retrieve before re-ranking.