FinSight writes financial reports where every claim carries a source

A financial report is more than text: its force comes from numbers you can check and charts that point back to their sources. LLMs write well but hallucinate. The FinSight team attacked that with a separate group of agents responsible for gathering data and verifying calculations, plus specialized vision-language models that make the final pass over how charts and tables are rendered. The result is a report the reader has neither the urge nor the need to re-check line by line.

What are we expecting at the output?

A structured report in which every section is a combined block: concise but coherent text, a visualization, and links to sources. The text carries the author's point, the visualization backs that point with data, and the links show where the data came from. The output should be a report that is both easy and worth reading — and above all, everything it says should be plainly supported by data and source links. That sounds simple stated out loud, but in practice it is hard to automate a process that takes in a pile of heterogeneous sources and data-handling code and puts out long text with no hallucinations, with the sources used spelled out and linked.

How it works inside

The foundation is the Code Agent with Variable Memory (CAVM) framework and a shared variable space that holds data, tools and agents at once. Every CAVM step is reasoning plus code: it generates code, runs it, and updates the variables and the history. That is what keeps CAVM from losing the thread even across a large number of dissimilar tasks.

The pipeline has three stages

• multi-step web search with source verification; the second pulls data from financial databases and APIs or scrapes web pages. Crucially, if the analysis stage turns out to be short of some data, a targeted collection run can be fired again — the system is not afraid to go back for what is missing.
• At the analysis stage a dedicated agent calls tools, uses code to clean and aggregate the data, and splits the result into compact chains of analysis (CoA). Visualizations are brought to their final state over several iterations: at each step a vision-language model critiques the visualization and the plotting code is corrected automatically.
• The final text is written in two stages. First come the compact chains of analysis, carrying explicit labels for the references and visualizations that will follow. Then a plan is built and the long, connected text is written section by section, with explicit pointers to the source and visualization identifiers established earlier. This holds hallucinations down and keeps links and figures exactly at the points in the long text where they belong.

Why it works

Because every action of every agent working on the report — from pulling a specific company's quarterly tables to drawing the charts — is expressed as code in a shared variable space. That leaves plenty of room to maneuver while forcing the agents to speak one language, and it removes the seams between search, analysis and final text. The two-stage narrative writing, in turn, fuses the scattered pieces of the analysis chains into a single account without losing the checkable source links.

How the authors tested that it works

Several aspects of the report had to be scored at once, across nine metrics grouped into three categories. First, factual accuracy and how tightly the text is bound to its sources, plus agreement between text and images. Second, how useful the resulting narrative is to an investor: density of verifiable facts, coverage of the key topics, depth of analysis. Third, the quality of the narrative itself — structural logic, professionalism of the language — and the quality of the visualizations. A reference report and the candidate versions under evaluation were handed to Gemini-2.5-Pro as judge, which returned a score and a short justification for each of the nine metrics. One practical question was treated separately: how the length of the long output text relates to all nine quality metrics.

Spoiler: longer is not better. Scores rose only when the long narrative rested on solid structure and links to verifiable sources.

What comes out at the end

FinSight was used to produce reports on individual industries and companies — from utility power and fusion to logistics, education and specific firms within those sectors. The system's center of gravity is the CAVM multi-agent framework, fed by a wide range of external sources.

The iterative approach to building visualizations and the two-stage writing of the long text are worth singling out. Together they yield a report with visualizations in which every sentence is carefully backed by a link to a checkable external fact, each figure is apt and legible, and the structure carries you from data to conclusions quickly.

In the experiments the system clearly beats the baselines on factual correctness, on how verifiable its claims are, and on the depth and usefulness of the final report. It is also a fine example of how to feed heterogeneous external data into advanced multi-agent frameworks and get the result you want.

The catch is that the input scenario has to supply enough of that heterogeneity in data and tasks — as it does in writing a financial report worth reading, where good prose alone is not the point.