Off meta: Are we tired of "one single product that solves everything" that every single AI product has became?
grep didn't try to also do what awk does, and jq and curl did exactly what they needed to do without wanting to become an OS (looking at you emacs), can we have that in the AI world? I hope/think we will, in a few years, once this century's iteration of FSF catches up.
jedberg 20 minutes ago [-]
The thing that AI is best at is summarizing vast quantities of information. That means the most natural thing for an AI to do is be "the one tool to rule them all".
The more information it has access to, the more useful the answer can be. But that also means that it can answer all the questions.
skeeter2020 13 minutes ago [-]
>> The thing that AI is best at is summarizing vast quantities of information
by definition a summary is the best at nothing though, and the mentality that the best way to rule is from a single summarized interpretation is both flawed and scary. It's not answering all questions; it's attempting to provide a single summation dramatically influenced by training. Go ahead and incorporate this into your balanced and multi-perspective decision-making process, but "one tool to rule them all" is not the same thing and definitely not what we're getting.
add-sub-mul-div 49 minutes ago [-]
I'm tired of endless LLM spam submissions from people who only use their accounts here to advertise and self promote.
Aperocky 45 minutes ago [-]
LLM submissions are no different from tech submission of yesterday. But most people used to build tools that does one thing well instead whatever the current meta is.
iLoveOncall 1 hours ago [-]
Developing and serving GenAI models is highly unprofitable, so, no, we're not going to have that in the AI world.
Either those model developers & providers package them in as many services as possible so that they can be somewhat profitable, or they die, and we don't have model developers & providers anymore.
Aperocky 53 minutes ago [-]
Well, the product here has nothing to do with serving GenAI models. It's now application territory.
And I prefer unix philosophy vs. the Copilot product approach.
xmprt 13 minutes ago [-]
> When a question touches restricted data — student PII, sensitive HR information — the agent doesn’t just refuse. It explains what it can’t access and proposes a safe reformulation. "I can’t show individual student names, but here’s the same analysis using anonymized IDs."
This part is scary. It implies that if I'm in a department that shouldn't have access to this data, the AI will still run the query for me and then do some post-processing to "anonymize" the data. This isn't how security is supposed to work... did we learn nothing from SQL injection?
thunfischbrot 2 minutes ago [-]
In the strongest interpretation of that it would offer only data which the user is allowed to access. Why do you assume that them implementing a feature to prevent PII being accessed that they then turn around and return data which the user is not supposed to access?
fishtoaster 31 minutes ago [-]
> This is the part that doesn’t demo well. ETL pipelines feeding into BigQuery from every operational system: Salesforce, Zendesk, and a dozen other internal tools. dbt transformations that normalize and document the data. Column-level descriptions for every table in the warehouse, because an AI agent that doesn’t know what a column means will write SQL that looks right and returns wrong numbers.
I'm glad they called this out. For the first half of this, I kept thinking: "Either your answers are confidently wrong or you've done a ton of prep work to let your AIs be effective BI analysts." Sounds like it's the latter, and they're well aware of it!
truelson 2 hours ago [-]
Just going to say it... no mention of handling the security aspects of this. Scary.
This is cool, I should say, but I would be really worried about the security aspects. Prompt injection here could be really painful.
32 minutes ago [-]
georgeburdell 1 hours ago [-]
The article mentions that there’s an identification process and that at least some data has access control. What were you expecting?
truelson 1 hours ago [-]
You're wiring up a number of critical systems... and prompt injection here could be really bad. I worry about such systems with a single point of contact
jedberg 19 minutes ago [-]
Reading through it, I didn't see any mention of write access. It looks like the agent is strictly read-only with access controls.
pwr1 53 minutes ago [-]
We tried something similar at a previous company — ended up with 3 different bots all answering slightly differently depending on which doc chunk they hit. The consistency problem is real.
Curious how you handle updates. Like if someone edits the source doc, does the bot just start returning different answers or is there a review step?
jgalt212 27 minutes ago [-]
She opens Slack, types a question to our internal agent: “Give me the names of all employees who have recently complain about my leadership”
Prior having such a product it was such a chore for her to track down all the people who may have objected (dispassionately or otherwise) to my plans, strategies, objectives, etc.
meryll_dindin 2 days ago [-]
We're a 30-person ed-tech company. I built a Slack bot that connects our data warehouse, 250k Google Drive files, support tickets, and codebase so anyone on the team can ask it a question and get a sourced answer back. The bot took two and a half weeks to build; the data infrastructure under it took two years. Wrote up the architecture, where trust breaks down, and what I'd build first if starting over.
bob1029 33 minutes ago [-]
> The bot took two and a half weeks to build; the data infrastructure under it took two years.
This is the key lesson that everyone needs to step back and pay attention to here. The data is still king. If you have a clean relational database that contains all of your enterprise's information, pointing a modern LLM (i.e., late 2025+) at it without any further guidance often yields very good outcomes. Outcomes that genuinely shocked me no fewer than 6 months ago.
I am finding that 100 tables exposed as 1 tool performs significantly better than 100 tables exposed as 10~100 tools. Any time you find yourself tempted to patch things with more system prompt tokens or additional tools, you should push yourself to solve things in the other ways. More targeted & detailed error feedback from existing tools often goes a lot further than additional lines of aggressively worded prose.
I think one big fat SQL database is probably getting close to the best possible way to organize everything for an agent to consume. I am not going to die on any specific vendor's hill, but SQL in general is such a competent solution to the problem of incrementally revealing the domain knowledge to the agent. You can even incrementalize the schema description process itself by way of the system tables. Intentionally not providing a schema description tool/document/prompt seems to perform better with the latest models than the other way around.
croemer 2 hours ago [-]
Is the result good? Is it useful?
And why does your comment say you're a 30-person company but the title says 60?
jedberg 18 minutes ago [-]
> And why does your comment say you're a 30-person company but the title says 60?
AI hallucination? :)
r1290 60 minutes ago [-]
I just did the exact same thing for my company. I didn’t do the sql lite approach for gdrive though just a direct search.
The one part that is still difficult is the data modeling and table level descriptions etc. Maybe you make an update to a table - remove a column, etc. The 3rd party systems all have their schemas defined but the data warehouse is a bit more loose. So solving that really helps. Did you just use dbt schema to describe tables and columns then sync that to your bot? How did you keep it updated? And end of the day - worth building or buying? Also how did you track costs? I let users choose their model - but have learned it can get expensive fast. As I can see there are a lot of providers trying to solve this one thing. That said the data warehouse aspect is the loosely defined area and I can see dbt or one of those players try to build something.
mannanj 45 minutes ago [-]
Hi. thanks for sharing. One thing I'd like to know is how often do you validate the answers? If a human gives an answer like the one the AI is giving for example, you'd probably expect a margin of error of like 1% of making a mistake. The AI though, is it 1% or less - and who's validating it? Are you trusting it more or less than a human?
oliver236 39 minutes ago [-]
data engineering is all you need.
everything else is smoke
all ai applications are smoke and will be obsolete in a year
do not be deceived
ricktdotorg 1 hours ago [-]
is it just me or was the scrollbar purposefully hidden on this site? in chrome on windows, i found it very jarring and user-hostile to NOT know how far along i was in reading the article.
i make a judgement call early on: is this worth my time? my whole article calculation algo was thrown off by this.
do not like.
maxothex 1 hours ago [-]
[dead]
mritchie712 1 hours ago [-]
> The data infrastructure underneath it took two years.
grep didn't try to also do what awk does, and jq and curl did exactly what they needed to do without wanting to become an OS (looking at you emacs), can we have that in the AI world? I hope/think we will, in a few years, once this century's iteration of FSF catches up.
The more information it has access to, the more useful the answer can be. But that also means that it can answer all the questions.
by definition a summary is the best at nothing though, and the mentality that the best way to rule is from a single summarized interpretation is both flawed and scary. It's not answering all questions; it's attempting to provide a single summation dramatically influenced by training. Go ahead and incorporate this into your balanced and multi-perspective decision-making process, but "one tool to rule them all" is not the same thing and definitely not what we're getting.
Either those model developers & providers package them in as many services as possible so that they can be somewhat profitable, or they die, and we don't have model developers & providers anymore.
And I prefer unix philosophy vs. the Copilot product approach.
This part is scary. It implies that if I'm in a department that shouldn't have access to this data, the AI will still run the query for me and then do some post-processing to "anonymize" the data. This isn't how security is supposed to work... did we learn nothing from SQL injection?
I'm glad they called this out. For the first half of this, I kept thinking: "Either your answers are confidently wrong or you've done a ton of prep work to let your AIs be effective BI analysts." Sounds like it's the latter, and they're well aware of it!
This is cool, I should say, but I would be really worried about the security aspects. Prompt injection here could be really painful.
Curious how you handle updates. Like if someone edits the source doc, does the bot just start returning different answers or is there a review step?
Prior having such a product it was such a chore for her to track down all the people who may have objected (dispassionately or otherwise) to my plans, strategies, objectives, etc.
This is the key lesson that everyone needs to step back and pay attention to here. The data is still king. If you have a clean relational database that contains all of your enterprise's information, pointing a modern LLM (i.e., late 2025+) at it without any further guidance often yields very good outcomes. Outcomes that genuinely shocked me no fewer than 6 months ago.
I am finding that 100 tables exposed as 1 tool performs significantly better than 100 tables exposed as 10~100 tools. Any time you find yourself tempted to patch things with more system prompt tokens or additional tools, you should push yourself to solve things in the other ways. More targeted & detailed error feedback from existing tools often goes a lot further than additional lines of aggressively worded prose.
I think one big fat SQL database is probably getting close to the best possible way to organize everything for an agent to consume. I am not going to die on any specific vendor's hill, but SQL in general is such a competent solution to the problem of incrementally revealing the domain knowledge to the agent. You can even incrementalize the schema description process itself by way of the system tables. Intentionally not providing a schema description tool/document/prompt seems to perform better with the latest models than the other way around.
And why does your comment say you're a 30-person company but the title says 60?
AI hallucination? :)
The one part that is still difficult is the data modeling and table level descriptions etc. Maybe you make an update to a table - remove a column, etc. The 3rd party systems all have their schemas defined but the data warehouse is a bit more loose. So solving that really helps. Did you just use dbt schema to describe tables and columns then sync that to your bot? How did you keep it updated? And end of the day - worth building or buying? Also how did you track costs? I let users choose their model - but have learned it can get expensive fast. As I can see there are a lot of providers trying to solve this one thing. That said the data warehouse aspect is the loosely defined area and I can see dbt or one of those players try to build something.
everything else is smoke
all ai applications are smoke and will be obsolete in a year
do not be deceived
i make a judgement call early on: is this worth my time? my whole article calculation algo was thrown off by this.
do not like.
yep, that's what Definite is for: https://www.definite.app/
All the data infra (datalake + ELT/ETL + dashboards) you need in 5 minutes.