The News, Explained
OpenAI’s V7 Go case study describes a system that connects repositories such as SharePoint and Google Drive, extracts entities, facts, attributes, metrics, and relationships, and organizes them in a “Context Graph.” An entity is an identifiable object such as a company, person, or fund, and the Context Graph stores links among those entities and facts. Each fact remains linked to a citation from the original document. Source
Building a graph from documents
V7 Go connects extracted facts to new or existing records in the Context Graph. Recurring workflows query the graph first, and when the needed information is missing, the system can use RAG to search the original documents. RAG retrieves relevant material first and supplies it to the model as context for generating an answer. Source
V7 cites private-equity deal screening and insurance underwriting as examples. In one demo flow, the system extracts financial information, transaction terms, and management information from a Confidential Information Memorandum (CIM), a document used in reviewing a deal, adds source citations to risk items, and uses those facts to prepare an investment memo. Source
Long-running work and MCP access
For longer-running tasks, recent conversation content stays in the model’s active context while older information is stored in the graph for retrieval when needed. Active context means information included directly in the model’s current input. V7 Go exposes graph collection and query functions through an MCP server. MCP is a standard interface for connecting AI systems to external data and tools. V7 says customers can use them from compatible clients such as ChatGPT or build workflows through Codex MCP. Source
Assigning different models to different tasks
V7 says it uses GPT-5.6 Luna for structured extraction and high-volume work, while GPT-5.6 Terra or Sol handle chat, Go Agent paths, and steps that need more reasoning or tool use. It has also started using GPT-6 Astra for especially difficult graph queries. Source
In its Context Graph benchmark, V7 reported that GPT-5.6 Sol reduced tool-call errors from 2.7% with GPT-5.5 to 0.2%. In a separate set of very difficult graph questions, it reported 89% accuracy for GPT-6 Astra and 78% for GPT-5.6 Sol. V7 also said GPT-5.6 Luna cut per-document cost by 78% compared with GPT-5.4 mini. Source
Customer results reported by V7
V7 reported a 21x improvement in deal-screening speed for one asset manager, reducing a one-day process to 15 minutes. Another financial-services team reduced review time from more than 100 hours to less than 10 hours and saved $12,000 in expert cost per job. In insurance, V7 said providing agents with prior claims and policy history reduced claims-processing errors by 13.5% compared with a manual baseline. Source
V7 says its model-selection process continuously benchmarks citation accuracy, extraction quality across document types, answer accuracy, instruction following, latency, cost, and real enterprise workflows. Source
OYOPICK’s Take
A Context Graph that retains original evidence connects faster answers with answers people can check again. Its value becomes clearer when a company can trace which changed document an earlier decision relied on. V7 describes change-triggered flags and renewed review as a longer-term goal; readers should not assume those actions already work automatically in every workflow.
The model figures concern different tests and tasks. Rather than combine document-cost savings and difficult-query accuracy into one performance score, compare extraction and citation accuracy alongside cost per usable result for recurring document types. If shared context helps find and refresh stale evidence, AI could become a more consistent collaborator in using organizational knowledge.