Show HN: Microsoft releases Flint, a visualization language for AI agents (microsoft.github.io)

2 points by chenglong-hn ↗ HN
Data visualizations are the bridge between user and data.

But building AI agents that can generate visualizations reliably can be very tricky:

- simple chart specs can be reliable, but generated charts are often of low quality due to reliance on system defaults; - complex chart specs with explicit details can produce good-looking charts, but they are verbose and agents can struggle with reliability

We figured out it is a limitation on the language issue (not just AI capability thing) -- current visualization languages are a bit too low-level for AI agents, requiring them to explicitly make visual decisions that are supposed to be handled by a good compiler. Flint is a visualization intermediate language to address this issue, allow AI agents to solve this last-mile human-agent interaction problem. It provides a simple semantic-type based specification, and contains a layout optimization engine that can produce good-looking charts (filled with derived low-level details) from simple high-level specs. The result is also very human understandable and adaptable. Flint powers data formulator for generating visualizations (another open source project from microsoft https://data-formulator.ai/).

Flint is available open source, and we built a MCP server that you can directly plug flint in your favorite agent app to play with data.

126 comments

[ 2.9 ms ] story [ 49.8 ms ] thread
> requiring them to explicitly make visual decisions that are supposed to be handled by a good compiler

Isnt graphviz there for the same reason?

Edit: I see it is using JSON as the declaration language, I am OK with llms being "good at json" but a syntax also consumable by humans it is not!

In fact, Json as a common language for human in visualization has been around for a while! The benefit of declarative grammar is that users can effective manipulate specs through UI (drag and drop, clicks).

Btw, Flint is intentionally designed to allow agent skip low-level params like scale, axe, zero, step size etc (which are extremely crucial for "GOOD-looking") and they are dynamically optimized by the compiler. So AI agents can have a easier time.

> Json as a common language for human in visualization has been around for a while

Plant, Mermaid, Graphviz are all declarative textual representations designed for human authoring, JSON is made for tools. Its not a criticism just a statement that if interop across agent and human was intended this is not the simplest option.

right, in fact many small models still struggle with following json, some new forms are also needed
Absolutely. This is DOA honestly and not really better than what we had before.
It compiles into Echarts, but echarts already has a JSON co figuration spec
It's more like a simple high-level spec to make it easier. The idea is that you don't have to fill position / axes details just to make the chart work. The compiler has a bit of magic of using semantic types to optimize what parameters will be set in ECharts.

In some composite chart examples, the good-looking echart spec is like 5x longer than the simple Flint one!

The charts are very nice, and I think the visualisation layer for LLMs is a very interesting problem.

I’ve been building https://smalldocs.org for this exact reason. It’s an office suite for AI agents - but my main use case is giving a cli based LLM the canvas to express itself - charts, mermaid diagrams, etc. I’ve extended it a bit further to be a format for all types of work so the agent can embed slides and spreadsheets in a document.

Sample document: https://smalldocs.org/blogs/what-is-a-smalldoc

Source: https://github.com/espressoplease/smalldocs

interesting how you don't discuss literally anything about the project actually posted and spam your thing. Not pointing you out, seen many other comments like this on HN but always felt a bit weird about them
In my experience, HN's excellent mods are always happy to hear about users who give you the ick by (for example) spamming their project 9 times in 2 days. They welcome feedback at hn@ycombinator.com.
Can you please stop? You're crossing the line into spamming HN and we're getting complaints.

It sounds like cool work, but you shouldn't be using HN in this way. Linking to your project occasionally in a relevant context is fine, but this should be a (small) minority of what you post to HN.

From https://news.ycombinator.com/newsguidelines.html: "Please don't use HN primarily for promotion. It's ok to post your own stuff part of the time, but the primary use of the site should be for curiosity."

This is cool to see from a research team. A few weeks ago I was exploring a similar idea with ntcharts, where a user or LLM can specify a chart in a Golang or JSON object...

and then that spec would be rendered either to a Bubble TUI via NTCharts or to HTML/SVG via ECharts. That Echarts HTML could be naturally served by a Golang http service.

But Flint goes much deeper with semantic layers and settings optimizations. Perhaps a NTChart, or whatever terminal chart, could be a rendering target? I'll add it to the list to explore...

https://github.com/NimbleMarkets/ntcharts/blob/spec/spec/REA...

This is fun! We started thinking it would just be an engineering task in the beginning, but doing a solid intermediate language turned out to be a research project (the paper will be out soon).

Also, I find NTChart very fun, maybe we should add NT chart to the list of compilation backend for Flint so it works in the library. Putting a reminder here: https://github.com/microsoft/flint-chart/issues/45

Is there a specific explanation about how this is better or different than vega itself? https://vega.github.io/vega/docs/specification/

My understanding is that Vega was already an expressive DSL for visualizations and its probably already well spread through LLM training data.

Vega was a high-level language in the past for human, but now they can be a bit too low-level for AI agents! AI agents have to write a lot of low-level params just to make charts looking good, and the result is that programs are hard to write reliably for AI agents.

Flint is a higher-level abstraction, with simpler much shorter spec, and the compiler derives low-level decisions so that charts are looking good.

So: flint lets agent write short program that achieving good looking charts that had to be done with lengthy program in the past.

I'm sorry, but as someone who creates data visualisation as a big part of my job, I wouldn't say the charts on the website look good. Most aren't awful either, but by no means are they an improvement over what I'd get by telling any coding agent to make a chart with Vega-Lite or Observable Plot, and probably worse than if I had some decent instructions/skills.

I don't quite get what the goal of this is other than abstracting away a little bit of the complexity at the expense of flexibility. To me, the promise of LLMs is the opposite, I can get flexibility and customisation without the cost of complexity.

Some composite charts are quite annoying to be generated well (like bullet, waterfall etc), their Vega-Lite equivalent can be quite long if just starting from scratch.

The intention here is that Flint is a simpler abstraction to get basic setups right and any followup edits can be done on top of the first compiled outputs (thus not limiting expressiveness). It also makes it easier for user to manipulate (like swapping axes, click to change something, which can be very hard if LLM generates a complex chart spec upfront).

But for many basic stuff your intuition is completely right.

That's fair, I generally make charts for publication, so I spend much more time and effort on the details. But I can understand this being useful for quick exploration for some people.

Generally speaking, I suggest anyone interested in learning to make charts get familiar with grammar of graphics [0] libraries like Vega-Lite, Observable Plot, ggplot2, Altair. There is a bit of a learning curve if you're used to selecting chart types like in Excel, but once it clicks, it gives you virtually unlimited choices in the kinds of charts you can make. And with ggplot or Observable Plot [1], it's about the same number of lines as something like Flint.

0: https://data.europa.eu/apps/data-visualisation-guide/why-you...

1: https://observablehq.com/@observablehq/plot-gallery

Grammar of graphics has been the foundation of a lot of stuff and definitely worth learning for everyone!

A challenge with GoG is that it assumes configurations as second-class stuff, which makes it quite difficult for users to deal with things like changing formatter, scale, annotations. Flint kinda want to hide this aways (so Flint sets them on behalf of the agent or the user). But yeah, GoG is still the foundation for expressivenss.

I strongly disagree ;-)

The paper's line of reasoning seems to continue the endless subjective loop of assuming your viz framework has the right abstractions & defaults , which the next person will rightfully disagree with for their slightly different eval set

We found in practice:

- LLM's generate charts fine

- LLM's tweak charts fine

- LLM's take user feedback to tweak them fine

In that sense, going higher-level for abstractions, as is being argued for here, is strictly worse: it's better to give controls so the LLM can go deep and customize

In practice, we found the choice of json config language X vs json config language Y to be pretty equivalent across different charting systems (vega, plotly, perspective, etc), LLM's do them all fine

The harder parts were deciding what a good chart is (model, reasoning, context), and opposite of this approach, giving lower-level facility for doing user change requests on tweaks, interactivity, and tricky in practice, when they have a lot of data on it.

You are absolutely right. But note that we are actually on the same point here.

This is exactly why this is an intermediate language designed to get 95% stuff right easily (for expressiveness and reliability purpose), while 5% of more advanced case where the agents need to revise chart for other purpose can be done easily on top of the compiled low-level spec (low in terms of Vega-Lite etc, not SVG). We are not really designing a higher abstraction to replace existing ones.

In the past, the split is like 50% good at first run for some common stuff, all other stuff requires agent-loop or user involvement.

Our goal is to make it easy for most case, not everything needs a full multi-round trip agentic workflow to solve. :)

We are kinda all advanced users in fact, for a lot of users, they are easily get confused with the first time result if that is not as good, and the interactivity cost / multi-round isn't an option.

I was wondering the same for vega-lite, which is relatively high level, declarative, and looks similar to their syntax.
This is pretty crazy, literally built something almost exactly like this for a project I'm working on (a local-first AI agent that does work on folders while you sleep). Basically going from JSON "Lego blocks" to full reports (including charting, though a subset of what Flint offers). And with post-generation validation and retry steps.

Functions extremely well and the result is a very clear (and consitent) human-readable "output layer." Cool idea, fun to see people converging on similar concepts in the space.

That's awesome!

I find that besides training better models, designing new language for agents is also a super viable paths to improve their performance!

> simple chart specs can be reliable, but generated charts are often of low quality due to reliance on system defaults; - complex chart specs with explicit details can produce good-looking charts, but they are verbose and agents can struggle with reliability

N of only a few of us working on an analytics agent, I don't think we've been finding this to be the case. We've been impressed with just how good LLMs (even smaller open weight models) are at using Python and R for visualization. Often any shortcomings go away if we iterate a bit to about ambiguity. Are there any threads of research that could better support this claim or highlight where issues might be?

A simpler spec can be used by a simpler agent. So, maybe that's the use-case here... use by smaller/cheaper agents that run in parallel as opposed to large models running one visualization at a time.

Or at least, maybe that's the idea?

IME, Claude and ChatGPT do just fine generating ggplot models, but extensive customization can get a bit hairy.

we are considering also reliability, interactivity besides expressiveness. Simpler spec with good expressiveness comes handy when you want the agent to be reliably for non-expert users and with small models.
There’s an emerging pattern in agentic systems and this project is a great example.

A deterministic layer like a compiler or generator of code with some kind of IR that the LLM generates and feeds it with.

I feel we will be seeing this more and more in the near future.

A well designed intermediary enables both validation and control over the output independent of the AI. This changes the interaction model between human and AI from delegation to collaboration.
also user interaction afterwards -- if can be frustrating if the only way the user can interact with the chart is to chat with the agent again (simple spec allows easy UI interaction!)
Yes, this has been the pattern for agentic systems since the beginning: permissive generation, that retries over and over until it gets the right size shape through the hole, and the input validates.
I'm all in on this idea. Every piece of agentic coding I have done in the last month has been via an intermediate representation. Iteration is done in the IR layer mainly. It's remarkable how close you can get to a deterministic coding output using this methodology.
"For AI agents". I understand why everything needs to be marketed in this way, but it's just ... an easy-to-generate language for expressing charts. That's impressive! That's useful.
Isn't this literally made for AI agents to be accessed through an MCP server? Seems to me the AI agents part of the marketing is important.
But why be exclusive? Why not "Chart language for computer programs to generate"?

I don't want to use an agent at all, but i wouldn't mind generating some charts with an easy-to-generate markup language...

But for that we already have mermaid.js (and its precursor Graphviz/dot).

The only reason to use this instead of existing, mature ones designed for humans is if you are an AI agent.

Mermaid looks terrible.

It is only better than nothing for the purpose of showing it to people whether produced by an LLM, by a human, or by both.

Your employees may just accept the internal slop, but at some point, you have to show your charts to your customer.

Doesn't that depend on the style that you apply to it?
No? I don't think there is a way to meaningfully change the layout for non trivial charts?
Why can’t AIs generate for the “existing, mature ones?” Like the other commenter said, I’m not sure I get the “this is totally for AI” marketing. Why can’t AI use the existing ones and why can’t humans use this?
its not marketing to you, its marketing to agents looking for tools to use
the design here has some constructs (i.e., semantic types) that AI can use better than human at generation time; and then the generated spec can be easy for user to edit since there is no need for hard-coded low-level parameters!
Graphviz and mermaid are a shitshow.

anything more than their own handpicked examples and you're better off using d3 or yfiles. layering, clustering, boundaries, rearranging are all basic needs for text to diagrams. None support it.

Both suck at being any good for rendering diagrams from readable structured text. There is a gap to be addressed.

I find it lacks some easy way to do alignments or grouping, which makes editing frustrating. Could be a good language re-design opportunity.
yes and that led me to a more fundamental question : is it even possible for an easy way to do fine grained adjustments. The finer granularity brings complexity. that makes it unreadable (especially for the untrained) and hence set aside and forgotten. The other way is to narrow down and focus on a well defined subset of problems.

either way the number of people willing or compelled to learn it will be tiny. and it hence becomes a niche, perfect for a long-term side project but with no real return.

my conclusion was to stop looking and use D3 or custom code.

That's the best option for now. But it can also be frustrating to ask AI to do small edits on D3 just to fix some idioms (like switching order etc) and they kept messing up with other stuff accidentally. Thus I still believe have a language with native representation for these diagram concepts would be helpful.

Good news is that AI do make new language a bit more accessible than before! If your agents can use it well and you can steer easily, it will naturally be good adoption.

I'd ask the question in reverse too. Why not let the AI agent use mermaid.js/graphviz/dot?

Presumably Microsoft thinks this one is better. Why? And why would that answer be any different for a human vs an LLM?

Because these are completely different requirements.
Which are what, exactly?
Careful, they'll spring something XML-derived on us.
for agent to generate, but also easy for human to edit (especially with UI) :)
also worth noting that Vega-Lite is literally just fucking that and AI already does good job with producing JSONs for it
I work with someone who did a lot of work with this to improve our ability to generate awesome visualizations with little thinking. It's a very powerful language but needs guardrails and guidance, particularly if you want end users to be able to produce consistent and standardized visualizations without knowing anything about it.
Also guidelines sent to the agent may or may not get ignored if they are just part of the context :(
Did you read the post? It directly answers why this is specifically for AI agents.
This can’t be said enough. “Good for Agents” just means self-documenting, obvious ergonomics, save defaults, succinct (or controllable) output, programable interfaces, … all of which support human users too!
The "curb cut effect", right? I'm sure there's a human-friendly interface to MCPs myself. I think we're reinventing application scripting that way.
And as per packages, built on top of existing charting libraries.
Thank you for summarizing it this way. All their flowery language (esp from OP) seems like long for "we figure out how the chart should look based on its data". From their page:

> Instead of requiring verbose low-level parameters such as scales, axes, spacing, and layout, the Flint compiler derives optimized chart settings from the data, semantic types, chart type, and encodings.

> Flint is built by Microsoft Research in collaboration with the IDEAS Lab, Renmin University of China.

Interesting to see a Chinese University collab for once.

OpenAI and Anthropic can use this chart library during the production of their IPO Prospectus so their promises of infinite future revenue is gloriously displayed!!
I don't really understand the point of this, I feel like LLMs have been able to one-shot matplotlib since GPT 3.5. I have extensively used LLMs to do data viz and haven't run into any problems. What is a specific instance where an agent struggles to generate a visualization and Flint solves it?
>Instead of requiring verbose low-level parameters such as scales, axes, spacing, and layout.

Ok, Microsoft is conflating two different things here: LLMs don't really care about code being low level and verbose, they can read things like Assembly and SPIR-V just fine: visualization is the real issue in that LLMs have no natural understanding of spatial composition through visual comparison because they literally "see" things differently than humans, so the way to get around that is provide them with "visualization" in code form that they can easily reason about and understand, so basically anything that's not deeply nested and has hidden states that they have to reason about.

Also, Flint being stringly typed in JSON is a decision that I don't think I agree with. Looking at the actual spec, this could have just been a normal, human usable TypeScript library, and it would have been 100x better. Using their own example (excuse the formatting):

type SemanticType = "Category" | "YearMonth" | "Profit";

type ChartType = "Heatmap" | "BarChart" | "LineChart" | "ScatterPlot"; // extend as needed

interface ChartEncodings { x: string; y: string; color?: string; size?: string; tooltip?: string; }

interface ChartProperties { colorScheme: string; [key: string]: unknown; // allow other optional properties }

interface ChartSpec { chartType: ChartType; encodings: ChartEncodings; chartProperties: ChartProperties; }

type SemanticTypes = Record<string, SemanticType>;

interface ChartConfig<TData = Record<string, unknown>> { data: TData; semantic_types: SemanticTypes; chart_spec: ChartSpec; }

// The actual typed object literal: const chartConfig: ChartConfig = { data: {}, // replace with your actual data shape/type semantic_types: { game: "Category", period: "YearMonth", newUsers: "Profit", }, chart_spec: { chartType: "Heatmap", encodings: { x: "period", y: "game", color: "newUsers", }, chartProperties: { colorScheme: "redblue", }, }, };

(comment deleted)
sadly, I think we are stuck with JSON as the most reliable way to get data / code in and out of an LLM (could be worse, could be YAML) … I’m interested in custom DSLs that improve LLM predictability and it is quite nice to see that even the Microsoft dinosaur “gets it” … see the Contacts example at https://slangify.org/examples which does VCARD to JCARD round tripping as a way to easily roll your own DSL
I felt conflicted as well, json is portable and easy to parse / validate and edit. But many models do still struggle. There are some stuff from functional programming might be worth bringing back here.
[delayed]
It's often comes with missing keys, use wrong value type (e.g., list over dictionaries). Mostly a small model issue and open source models, they don't follow instructions on the structure guidance that well, and there is no easy way to do generation-time validation.
This happens if a company has a CEO who presumably can no longer successfully go to the toilet without AI assistance.

Agents, npm, typescript, MCP. All buzzwords are there. Will anyone look at the slop charts? Of course not, the tokens are the goal.

MSFT stock is at 2024 levels. Maybe someone should produce a flint chart and present the agentic work to Nadella. No one buys this AI slop any more.

I actually made a chart with Flint to show MSFT stock, and with sparkline chart to compare with other companies... :)
Nice to stumble over this thread.

I'm not sure if Flint is the right tool for me. I'd like to have a tool that expresses code in visual form for me. For example, right now I need to reverse engineer some code for debugging purposes.

I already found out there are three tasks:

    * Task one fills task two's queue and waits for an event to get notified
    * Task two reads from its queue, forwards elements to task three's queue.
    * Task three reads from its queue and sends a success/fail message back to task two's queue
    * Task two then notifies the waiting task one.
Visually it's easily expressed: 3 bubbles lined up with 2 connections between the neighboring ones.

Which ML tools suited best for that?

Seems a bit more like https://mermaid.js.org use cases!
Here's the official mermaid vscode ext: MermaidChart.vscode-mermaid-chart: https://marketplace.visualstudio.com/items?itemName=MermaidC... .. src: https://github.com/Mermaid-Chart/vscode-mermaid-chart :

> @mermaid-chart

The functionality of the vscode-mermAId extension was merged into vscode FWIU? From https://code.visualstudio.com/updates/v1_109#_mermaid-diagra... :

> Mermaid diagrams in chat responses

> Chat responses can now render interactive Mermaid diagrams with the renderMermaidDiagram tool. This lets models use flowcharts, sequence diagrams, and other visualizations to visually break down complex concepts. The diagrams are interactive, so you can pan and zoom to explore them in detail, or open them in a full-sized editor for easier viewing.

Probability of an MS project existing in the next N days:
This is a valuable method of closing the gap in making LLM results available in a good visual form. The idea of viewing the charting process as the last step—where the AI deals only with high-level semantic specifications as opposed to all the low-level visual details—makes so much sense. I'm interested to know what the layout optimization engine does: can the developers inject the desired limitations for some cases, or is the layout design process a black box? It's good to see you've made it open-source!
Flint: because watching an AI agent work is like watching a Roomba navigate — you need a visualization to understand why it spent 20 minutes in the corner
more like, instead of watching the AI agent 20 minutes to create a polished chart with 100k tokens, Flint lets you get there in 20 second with 200 tokens with a 95% polished chart. :)