From 24% to 83% — What Actually Changed
Snowflake unveiled Cortex Sense at its Summit on June 2, 2026. It is a runtime layer that automatically gathers data and activity signals and enriches, in real time, the context an agent will reference at the moment a query arrives. The number it put on stage was accuracy climbing from 24% to 83% on the same query set — and what moved that needle was not the base model or the embeddings, but the context assembled at the instant of the query.
Two Layers: Governance and Runtime
The same announcement paired Cortex Sense with Horizon Context, the layer where a governance team curates business semantics — the definitions of metrics and terms. Definitions are governed once in Horizon Context and deployed everywhere; Cortex Sense then pulls and assembles them at runtime. The design turns on that split: humans fix meaning through semantic curation, and the machine fills it in at query time through automatic enrichment.
The First Vendors to Put a Number on the Bottleneck
What makes this release notable is that a vendor, for the first time, put a number behind the claim that agent accuracy is bottlenecked not by the model or the retriever but by context assembly. Databricks points the same way, claiming Genie answers correctly on the first attempt 84.5% of the time. The signal is plain: before you tune the retriever further, fix which definitions get assembled at query time.
Before You Tune the Retriever: A Two-Stage Mini Context Layer
Even a small team that cannot buy both layers wholesale can nail down the metrics first. Make first-attempt accuracy the primary KPI — the share of questions answered correctly with no re-prompt — and set a quarterly target of lifting it from below 40% to 80% or higher. The second metric, context hit rate, is the share of responses that actually included the definition they needed to reference; treat 90% as the pass bar. The third is definition-conflict count: tally the terms used differently across departments and drive that number toward zero, which is the governance layer's job.
Failure tends to arrive in three forms. The first is swapping embedding models over and over. Retrieval ranking may wobble a little, but if "monthly revenue" means different things to different teams, accuracy stays flat. The second is letting cross-department terminology drift go unaddressed: when finance's "active customer" and marketing's "active customer" point at different conditions, the same question returns different numbers. The third is having no owner for a semantic definition — if nobody is on the hook to fix and change a definition, the curation Horizon Context describes never gets off the ground.
Recovery starts by predefining the branch to take when a metric leaves its threshold. When context hit rate slips below 90%, fill the missing definitions into the dictionary table before you touch the embeddings; when a new conflict surfaces, hand the term to its owner and collapse it into a single definition. Queries whose first-attempt accuracy regressed get pulled from the logs and auto-enqueued onto the next curation backlog.
You do not need a heavyweight platform: a mini context layer begins as two spreadsheets. One is the dictionary table — columns for term, agreed definition, formula or SQL, owner, and last-updated date, with only cross-department-agreed entries admitted. The other is the query log, capturing the raw question, the definitions used in assembly, the first-attempt outcome, and a conflict note, so the three metrics can be tallied weekly. In the agent prompt, inject the dictionary table's agreed definitions at query time — that is the two-stage structure in miniature, letting the runtime use exactly the meaning a human governed.
Lock quality control to a pre-deployment scenario check. Bundle 20 representative questions per department into a golden set, re-measure the three metrics every time the dictionary table changes, and scrub personally identifiable information out of the definition columns so no sensitive data leaks in.
Improvement runs as a weekly pass: pull the top questions with low first-attempt accuracy from the query log and ask owners to strengthen those definitions. Stamp each newly agreed definition with an update date so stale entries get re-reviewed each quarter, and read any renewed rise in definition-conflict count as a sign that curation ownership has blurred. Once this loop turns, context hit rate climbs before you ever swap the retriever.
Takeaways at a Glance
What 24%→83% points at is context assembled at query time, not model horsepower. Split human-governed curation from runtime auto-enrichment into two layers, fix the targets — 80% first-attempt accuracy, 90% context hit rate, zero definition conflicts — and run the mini layer on nothing more than a spreadsheet dictionary table and a query log; the accuracy bottleneck loosens before you re-embed a thing.
References
Snowflake Summit 2026: Context, custom model training, Iceberg V3 — Constellation Research
Snowflake Cortex Sense and the Enterprise Context Layer — Atlan