Fluree AI

This is why AI hallucinates. Messy data in, and sometimes even garbage out. I love that someone is actually solving this at the source. Congratulations on your launch! :) How does this work for someone who is not a data expert but needs their AI to stop making things up?

Hi @bplatz @kevin_doubleday Great job and I played with the product. is there a way to connect with my claude code and run this? Honestly as much as I love it. It will take me a time to fully get the potential of what you have build and is there guide I can use. Btw love the way you have thought of the entire KG

the trust part is the harder half of this problem imo, not the ingestion. once an agent has already pulled a piece of context into its working memory and acted on it, and you later discover that fact was wrong or outdated in the graph, is there any mechanism to claw that back or flag the agent's prior output as suspect? or is it on the app builder to re-verify after every graph update?

Love the focus on verifiable AI instead of just faster AI. How do you handle conflicting data sources? Congrats on the launch! 🚀

I really like that you led with trust and permissions rather than just speed. That is the part most teams bolt on later and end up regretting, and putting it first says a lot about how you think about this. From my own time trying to get dependable answers out of messy data, the moment that always wins people over is when a tool can simply say "I do not have that" instead of guessing. If Fluree does that gracefully, you are going to earn a lot of trust quickly. Congrats on the launch.

Congrats on the launch. The detail that stands out is checking permissions on every request instead of only at ingest. That is the failure most teams find late, when an agent summarizes something into an answer the asker was never cleared to see.One builder question: when the graph auto classifies data and gets a relationship wrong, how visible is that to the person asking? Citations help a lot, but a confidently wrong join is harder to catch than a hallucinated sentence.Good call on the MCP side, that is the right seam for this.

Most data is in data lakes at other orgs.

Say I have most of my data in snowflake, can fluree connect to other warehouses/lakes?

TEST DO NOT POSTThe part I like here is treating trust as a data-layer problem instead of a prompt-engineering problem. Once agents start touching real operations, provenance and permissions stop being enterprise checkboxes and become the difference between useful automation and expensive guesses.

It feels like we’re getting past the stage of building on top of models and starting to solve the harder problems around trusted data and context.

A graph-based approach to memory and verifiable info makes a lot of sense.

Looking forward to seeing where this goes.

Congrats on the launch 🚀
@bplatz @kevin_doubleday

The part I like here is treating trust as a data-layer problem instead of a prompt-engineering problem. Once agents start touching real operations, provenance and permissions stop being enterprise checkboxes and become the difference between useful automation and expensive guesses.

The part I like here is treating trust as a data-layer problem instead of a prompt-engineering problem. Once agents start touching real operations, provenance and permissions stop being enterprise checkboxes and become the difference between useful automation and expensive guesses.

Interesting launch! Especially - building permissions into the query layer instead of bolting them onto the model, that's the exact spot access control leaks when it's an afterthought IMO. Do you guys surface the actual query that ran so I can catch an intent miss myself?

Hey. So If the data is correct, the output is correct" is a clean line but it quietly assumes correctness is binary, when in most real datasets it's not. A CSV can be internally consistent and still be stale, or two systems can both be technically correct about the same customer using different definitions of an active account. Structured queries against a graph solve hallucination, but they don't solve disagreement between sources about what the ground truth actually is. How does Fluree handle it when two ingested sources genuinely conflict rather than one just being wrong, since that is a different problem than an LLM making something up.
Also on the plug in any AI part, if Claude, OpenAI, Gemini, and Ollama can all reason over the same graph through MCP, does the cited and reproducible guarantee hold equally well across all of them, or does the quality of the citation depend on how well a given model respects the structured query results versus improvising around them.

The strongest part here is putting provenance into the context layer instead of leaving it for dashboards after the fact. One edge case I would want to understand: when two source records conflict, does Fluree expose both to the agent/user, or resolve them into one trusted context before the model sees it?

the "throw everything at it and the graph builds itself" part is what I keep coming back to. auto-classifying data and mapping it to an ontology is exactly the kind of thing that works great 95% of the time and quietly gets one relationship wrong the other 5%. when the automatic mapping misclassifies a field or links two different concepts together, is there a review step before that goes live, or does it just start feeding agents until someone notices an answer looks off?