V7 has turned to GPT-5.6 Luna to power a new approach for organizing and extracting knowledge from large, scattered document sets, and reports dramatic gains in speed, cost and accuracy. The company’s V7 Go platform ingests files from common corporate repositories, indexes entities and facts into a structured Context Graph, and orchestrates long, document-heavy workflows so teams can complete multi-step tasks in minutes with source-linked evidence.
At the heart of V7 Go is the Context Graph, which maps entities such as companies, funds and people to the facts and metrics found across an organization’s documents. The graph links each extracted fact back to its original source, preserving citations so agents can retrieve context without re-searching multiple systems. When the graph lacks sufficient context for a query, V7 Go falls back to retrieval-augmented generation against the underlying documents; for longer-running agent processes, recent exchanges remain in the model’s active context while older material is archived in the graph for on-demand retrieval.
V7 reports that agents using the Context Graph can complete 50–100 step workflows in minutes and achieve 99.9% accuracy, with an auditable trail for every decision. Measured customer outcomes cited by V7 include screening deals 21× faster, shortening a financial review from more than 100 hours to under 10 hours (saving roughly $12,000 in expert costs per task), and reducing errors in claims processing by 13.5% compared with a manual baseline. For high-volume extraction work, the company says moving those workloads to GPT-5.6 Luna yielded a 78% lower cost per document relative to GPT-5.4 mini.
V7’s testing also highlights model-level performance differences across its workloads. For tool-calling within the Context Graph benchmark, GPT-5.6 Sol reduced errors from 2.7% with GPT-5.5 to 0.2%. In a tougher graph-query suite using messy, real-world data across thousands of documents, GPT-5.6 Sol scored 78% accuracy on the test’s very-hard level while GPT-6 Astra reached 89%; both models performed near 100% on easy, medium and hard levels. The company additionally reports moving document-heavy workloads from the Chat Completions API to the Responses API, observing roughly a 5% token reduction for some PDF-heavy workflows and improved caching reliability.
Operationally, V7 assigns roles across models to balance cost and capability. GPT-5.6 Luna is tasked with structured extraction and high-volume jobs, while GPT-5.6 Terra and Sol handle chat, agentic decision paths and steps requiring more reasoning or tool use. For the most demanding Context Graph queries, V7 has begun testing GPT-6 Astra. V7 Go itself orchestrates long-horizon workflows that combine deterministic code, handoffs to smaller models, document generation and external integrations, and it maintains an auditable trace of every run to help teams manage risk and verify outputs for mission-critical processes in finance, insurance and real estate.
Beyond raw model selection, V7 emphasizes developer experience and integrations. The Context Graph is exposed through V7’s MCP server so customers can query and ingest context via ChatGPT and other compatible clients and build V7 Go workflows in Codex. V7 says streamlining workflow design has reduced the time to build a medium-length workflow from about an hour to roughly 20 minutes. The firm also runs a continuous benchmark suite covering citation accuracy, extraction quality across hundreds of document types, instruction following, latency and cost, and reports that OpenAI models outperformed other providers on prioritized behaviors.
V7 positions its approach as a practical answer for document-heavy enterprises that require reliable retrieval and auditable outputs. By organizing buried context into a structured graph and assigning roles to different models, the company aims to reduce repeated searches, lower token consumption and limit hallucinations on unanswerable queries. In internal tests, V7’s retrieval-only system outperformed a HERB baseline by 69% and cut hallucinations on unanswerable queries by 38%. Co-founders framed the effort as teaching AI a company’s operational knowledge as well as it has learned from the internet, and noted that stronger models and richer context have shortened delivery timelines and removed intermediate workflow steps.
V7’s results—especially the reported 78% reduction in cost per document with GPT-5.6 Luna and the dramatic workflow speedups—underscore the potential operational and financial impact of combining specialized models with a structured knowledge graph. For organizations wrestling with large, disparate document stores, V7 Go presents a tightly integrated stack that emphasizes traceability, efficiency and measurable gains in both accuracy and cost control.
Source: Read the original source
Leave a Reply