Motivation

• Speed for query and search tools
• Improved metadata
• Laser-focused context

Speed

• ClickHouse columnar database designed for real-time analytical queries on massive datasets with exceptional speed
• Qdrant vector database designed for storing, searching, and managing high-dimensional vector embeddings with filtering capabilities

https://console.cloud.google.com/compute/instances?authuser=5&project=chat-plot-staging

Better Metadata: Data Catalog

https://github.com/Roker-Research/data-catalog

• Current workflow
  • Gram starts from scratch every time
  • Northstar starts from scratch every time
• Proposed workflow
  • Constant, incremental improvement
  • Human-guided surgical updates
  • Fewer tokens
  • Fast ingestion

Better Context

https://staging.axes.com/admin/system-prompt

https://staging.axes.com/admin/search

• Variables vs Columns
• Seeding the context with dataset list
• Examples in system prompt
• Clean output

Advantages

• Ingestion ~10 minutes for 3 datasets— No LLMs!
• Faster tool calls Faster Chats
• Fewer tool calls Cheaper
• Cleaner Context Better Agents (also faster and cheaper)
• Lower server costs (ClickHouse $97/month, Qdrant $6/month)