Volcengine ContextSearch raised the business correctness rate of SQL generated by Query Agent from 36.8% to 98.7% across a test set of 76 advertising-business queries. We believe these results matter more than “another Agent product”: enterprise AI’s weakness is shifting from generation capabilities to the ability to find and correctly use knowledge inside an organization.
The key change is not that the model writes more complex SQL, but that before generating a query, it verifies how “active campaigns” are defined, whether actual spend excludes refunds, and how tied amounts are handled in the business documents and code repository. In other words, the model is not asked to fill in missing logic through speculation; it verifies first, then executes.
What this is
ContextSearch is an intelligent search service for AI Agents. It connects business documents and code repositories, organizes statistical definitions, field meanings, and value rules into a unified semantic layer, and makes them available to Query Agent through Agentic Search—a mechanism that lets Agents autonomously retrieve information, supplement it, and correct queries around a task.
Take “the active ad campaign with the highest actual daily spend” as an example. The system first decomposes the query into three implicit conditions: how to identify an active campaign, how to calculate actual spend, and how to select a result in the event of a tie. It then generates read-only SQL and checks the execution results. When information is insufficient, it should explicitly identify the gap rather than fabricate a plausible-looking answer.
This case comes from Volcengine’s internally built advertising daily-report dataset. The sample contains 76 questions, and human reviewers validated the filter conditions, statistical methods, and execution results. It demonstrates an improvement in a specific scenario, but it cannot yet be treated as a general performance level across industries and data environments.
Industry view
The positive signal is clear: enterprises do not necessarily need business teams to manually build a complex knowledge base first. Once they declare the sources of documents and code, the system can extract and continuously maintain some business knowledge. In this case, tool calls for the first question fell from 19 to 4, cutting execution time by 85%. That suggests checking rules before generation is more efficient than having an Agent repeatedly try SQL queries.
We also maintain three cautions. First, a test scale of 98.7% is not large and covers only advertising-campaign daily reports. Second, code and documentation may not agree; the system has not disclosed enough about how it handles conflicting information, outdated rules, or applicable versions. Third, automatic knowledge extraction may still misinterpret business definitions. If business personnel cannot review and revise the output, the errors merely move from the SQL layer to the knowledge layer.
ContextSearch is therefore better viewed as a potential infrastructure layer for enterprise data queries, not a fully reliable “automatic answer” for business questions. Final responsibility still rests with joint governance by data, engineering, and business teams.
Impact on regular people
For enterprise IT: AI project evaluation will expand from model selection to internal knowledge integration, access control, and version governance. The more complete an organization’s existing documentation and code, the lower its deployment cost may be; systems with fragmented knowledge will amplify erroneous decisions.
For individual careers: Natural-language data queries from operations, sales, and management roles may receive faster responses, but that does not eliminate the need to verify data. The ability to define business metrics, validate results, and feed corrections back into the system will become more valuable.
For the consumer market: The near-term impact is limited, as gains in enterprise efficiency will not immediately translate into more mature consumer Agents. However, as data-query, customer-service, and after-sales systems increasingly connect to business knowledge, users may encounter fewer instances of “the feature works, but the answer is wrong.”