A maintained securities context.
Companies, filings, financial periods, statements, notes, entities, and exhibits are structured into a continuously updated research system.
Unlocking machine intelligence for capital markets.
General-purpose models understand language. They were not built with a continuously maintained understanding of the public-company record.
VEKSTAR-01
Financial intelligence must begin with current, structured evidence.
Useful analysis requires more than access to a document. The system must identify the right issuer, form, period, statement, table, disclosure, relationship, and source passage—then determine whether the evidence supports the conclusion.
Reasoning governed by the record.
The research process begins by assembling the authoritative evidence required for the question rather than relying on broad model memory.
Verification as a delivery state.
A conclusion receives verified status only when the retrieved record supports it under the system’s review criteria.
Financial question answering
Benchmarking the work general-purpose systems repeatedly miss.
FinanceBench tests open-book financial question answering against source documents and supporting evidence. Vekstar’s evaluation framework focuses on retrieval, calculations, cross-document reasoning, source support, and abstention.
FinanceBench accuracy
Prototype result supplied for this website concept. Publish only with a complete, reproducible methodology and answer-level results.
GPT-4 Turbo with retrieval
The original FinanceBench paper reported that the evaluated GPT-4 Turbo retrieval setup answered correctly on 19% of the tested questions.
*Before public launch, replace this prototype claim with a linked report disclosing the evaluated question set, model version, retrieval configuration, prompts, calculation treatment, abstentions, scoring rules, and reproducible results.
Evaluation standard
Every performance claim should be as reviewable as every product answer.
Define the corpus.
State which filings, versions, tables, and supporting documents were available to the system.
Publish the configuration.
Disclose models, retrieval, prompts, tools, calculation methods, and verification thresholds.
Show the errors.
Release answer-level outcomes, failure categories, abstentions, and known limitations.
Technical reports
A research program built around evidence, not demos alone.
| Report | Scope | Status | Action |
|---|---|---|---|
| VEKSTAR-01 System Card | Architecture, source controls, verification, limitations, and deployment. | Draft | Publish PDF |
| FinanceBench Evaluation | Reproducible benchmark methodology and answer-level results. | Required before claim | Publish methodology |
| SEC Corpus Coverage | Document counts, issuer coverage, form coverage, latency, and exceptions. | In validation | Publish data sheet |
| Verification Gate | Support criteria, contradiction handling, confidence, and abstention. | In development | Publish technical note |
The goal is not to make AI sound certain. It is to make every conclusion inspectable.Explore a live workflow
See Vekstar on the companies you work on.
We will configure your demonstration around the companies, filing types, peer sets, and highest-cost research workflows your team already covers.