All AI videos

AI video

Why Your Company Needs a Context Graph (and How to Build It) — Gil Feig, Merge

Merge co-founder Gil Feig describes a context layer that routes between synchronized search, live API calls, business skills, memory, and source provenance.

AI Engineer · 2026 featured video

Watch on YouTube

What this video shows

Feig uses customer sentiment questions to separate direct lookups from searches across large historical datasets. A live API can retrieve one known record, while a synchronized index supports filtering and semantic retrieval across many records. A router chooses the source, and a summarizer combines the evidence with company-specific skills and memories.

Learnetto finds the live-versus-synced decision useful, but the term context graph is broader than the graph database definition in this talk. The speaker presents a vendor architecture rather than comparative evidence. Teams should model permissions, freshness, deletion, provenance, and source conflicts before choosing storage or integration products.

Airbyte documents a similar distinction in its Context Store and live connector model, including OAuth and searchable replicas. Use the MCP architecture documentation to understand what the protocol provides before adding a separate context layer.

What you will learn

  • Use live calls for current known records and synchronized search for broad historical or semantic questions.
  • Store the source, retrieval time, permissions, and invalidation conditions with derived context.
  • Express company definitions as reviewed skills or rules so users can inspect what terms such as active customer mean.
  • A router needs evaluation cases for stale copies, missing permissions, conflicting records, and unsupported queries.

How to apply this safely

  1. List the top agent questions and classify each as direct lookup, filtered search, semantic search, or derived analysis.
  2. Choose the authoritative source for each field and specify synchronization, deletion, and permission behavior.
  3. Return citations and timestamps with every retrieved record or generated summary.
  4. Test known-record lookups, corpus-wide questions, revoked access, stale data, and disagreement between live and synchronized sources.

Important limitations

  • The speaker is a Merge co-founder, and the final architecture supports Merge products. The talk does not compare retrieval quality, latency, or cost with other systems.
  • A synchronized copy creates security, retention, residency, and deletion obligations. The video gives a conceptual architecture rather than a complete governance design.

Sources to check

Continue learning on Learnetto

AI evals guide

Test retrieval, permissions, freshness, and answer grounding.