Beyond RAG: Build a Relational Context Engine from Scratch — Peter Werry, Unblocked

Similarity search fails on temporal or structural questions like 'PRs I merged last week'. Having the model write a validated database query, with allow-listed operators and tenant isolation, fixes this safely.

• Schema discovery samples 300 documents and merges them into one tree with procedural code rather than a model, giving the query planner reliable field types and enums.
• Identity resolution uses simple fuzzy matching to map a name like Peter to a GitHub login, and the query planner receives the schema, a forced name-extraction tool call and the current date, which questions like "last week" depend on.
• Validation is deny by default: an operator allow list stops a function clause from running JavaScript on the server, and hidden metadata fields are wiped and reinjected to keep tenants isolated.
• A retry loop feeds validation and execution errors back to the model, and those failures exposed the need for full-text search where exact matching missed.

• context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
• embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.

• videoAI Security Engineer Foundations + Certificate — Micah Silverman, Snyk
• videoThe SDLC You Build Yourself — Andy Wong, Andrei Bocan & Shane Wolf, Atlassian
• videoAgent Speedrun — Elizabeth Fuentes Leone & Sandhya Subramani, AWS
• videoCan Your Agent Hear You Now? Building Live Voice Agents with Gemini — Thor Schaeff

• videoBeyond RAG: A Relational Context Engine That Cuts Token Burn — UnblockedAI Engineer
• videoHow to Generate Mergeable Code with a Context Engine — Peter Werry, UnblockedAI Engineer
• videoYour agents lack context: Here's how to fix "You're absolutely right!" — Brandon Waselnuk, UnblockedAI Engineer

Checking sign-in…

Loading comments…