Context Manager

When to Use This Skill

• Designing RAG (Retrieval-Augmented Generation) architectures
• Optimizing context windows and token budgets
• Orchestrating multi-agent context handoffs
• Designing Vector Database schemas (Pinecone, Qdrant)
• Building Knowledge Graphs for semantic reasoning
• Implementing intelligent memory (short vs long term)

Workflow

1. Analyze: Determine scope (User Session, Project Lifetime, Enterprise).
2. Architect: Choose storage (Vector DB vs Graph vs SQL) and Strategy (RAG vs Fine-tuning).
3. Optimize: Implement chunking, ranking, and compression strategies.
4. Orchestrate: Define how agents share and update state.

Instructions

1. RAG Strategy (Retrieval-Augmented Generation)

• Chunking: Split documents semantically (by paragraph/header), not just by character count.
• Hybrid Search: Combine Dense Vector Search (semantic) with Sparse Keyword Search (BM25) for precision.
• Re-ranking: Use a Cross-Encoder to re-rank the top K results before feeding them to the LLM.

2. Context Window Optimization

• Compression: Summarize older turns in a conversation.
• Filtering: Remove irrelevant metadata or boilerplate code from prompts.
• Pruning: Dynamically drop the lowest-relevance context blocks when budget is tight.

3. Intelligent Memory Systems

• Episodic Memory: "What did we discuss 5 minutes ago?" (Recent chat history).
• Semantic Memory: "What are the user's preferences?" (Long-term facts stored in Vector DB).
• Procedural Memory: "How do I perform this task?" (Stored skills/workflows).

4. Knowledge Graphs

• Entities: Nodes (User, Product, Order).
• Edges: Relationships (User -> Purchased -> Product).
• Reasoning: "Find all products purchased by users who also bought X".

Resources

• Advanced RAG Patterns
• Context Optimization
• Multi-Agent State