Cerebras Designs a RAG-Based Internal Knowledge Base

Cerebras built the Cerebras Knowledge system, which integrates fragmented internal data from Slack, GitHub, Google Docs, and more to answer questions in natural language. This system processes over 15,000 questions daily, and it adopts a sophisticated design that goes beyond simple vector search to secure practical accuracy.

The key design features are as follows:

• Data Ingestion Strategy: Instead of forcing changes to user behavior, the system extracts data directly from existing tools (Slack, Jira, etc.), maintaining continuity of data.
• Unified Datastore: Built on PostgreSQL, it stores embeddings, original summaries, and metadata in a single table, allowing data from various sources to be queried through the same interface.
• Search Optimization: To overcome the limitations of simple embedding search, it combines vector search using pgvector with reranking via RRF (Reciprocal Rank Fusion) to improve search quality.
• Extensibility: The system is designed so developers can build custom connectors directly, allowing the system to be flexibly expanded as new data sources are added.