How query logs fix AI agent SQL errors

Justy Sixty-five percent. That's the number that stuck with me.

Cody Miro pointing agents at Snowflake and getting it wrong two out of three times?

Justy Yeah — and it wasn't the model, it was the context. Ten thousand tables, no semantic layer, and the agents just… hallucinated joins.

Cody Right.

Justy Which is wild because that's the exact pitch everyone's making for text-to-SQL right now. Just point the agent at the warehouse!

Cody And it falls apart the second your org is big enough to have multiple teams naming things differently. Schema tells you what columns exist. It doesn't tell you which ones actually go together for a business question.

Justy Okay so DataHub's new thing — Context Intelligence — basically says, instead of giving the agent raw schema, give it the query history. The joins that already worked.

Cody That's the core argument, yeah. They've been doing lineage tracking for years, pulling query logs from warehouses. Now they're flipping that same infrastructure to build what they call a semantic index.

Justy Wait, so the plumbing isn't new?

Cody That's the part I actually like. They didn't build this from scratch. The query log extraction and SQL parsing was already running in production for lineage. Postgres, Snowflake, BigQuery — they've been pulling from those for years. The new layer sits on top.

Justy That's smart product sequencing. You already have the pipes, now you monetize the insight.

Cody Sure, but let's talk about what they're actually doing with those logs. They filter for what they call golden queries — high-quality analyst queries and scheduled pipelines that represent proven business logic.

Justy Mm-hm.

Cody Then they invert the SQL patterns into structured text definitions they call semantic anchors. Those anchors become the retrieval basis the agent queries before it generates SQL.

Justy So instead of the agent guessing how to join tables, it looks up what joins humans already validated?

Cody In theory, yes. And there's a human review step on top — domain experts can resolve conflicting definitions and simulate changes before publishing.

Justy Cody, I have to ask — how many orgs actually have clean, consistent query history? Like, is the golden query problem itself a mess?

Cody That's exactly where I push back. If your analysts are writing garbage SQL and your scheduled pipelines are held together by duct tape, your golden queries are just well-documented garbage.

Cody The approach assumes a certain maturity. Orgs that actually have analysts who know the data and write consistent queries — this helps them a lot. Orgs where everyone's just winging it? Query history won't save you.

Justy But for the Miro case — ten thousand tables — someone's been writing correct queries against that environment. You're not starting from zero.

Cody Right, and that's where I think the argument does hold. If you're a company big enough to have that many tables, you almost certainly have analysts who know the right joins. The problem is that knowledge is trapped in their heads and in their query tabs.

Justy Extracting it is the actual product insight. Like, the reframing here is — this is a retrieval problem, not a generation problem. The model doesn't need to be smarter about joins. It needs better context about what's already worked.

Cody I agree with that framing. The model was never going to figure out that your finance team calls it gross_revenue and your sales team calls it total_bookings from first principles. That's lookup, not reasoning.

Justy And they're exposing this via MCP, LangChain, Google's Agent Development Kit, and CrewAI. So it's pluggable.

Cody Which is table stakes for this kind of thing honestly. If you built a context layer and only exposed it through your own SDK, nobody would adopt it.

Justy True. Okay, I got sidetracked earlier — I was looking at my coffee and realized I haven't actually had coffee yet today. I got up, started reading this DataHub piece, and just… forgot.

Cody That is very on-brand for you.

Justy It really is. Anyway — who do you think this actually changes things for? Like, practically?

Cody Data teams at mid-to-large companies who've been putting off building a semantic layer. This is basically a shortcut to one, derived from behavior instead of declared rules.

Justy That's a way better pitch than most semantic layer vendors give. They all say 'define your metrics once.' This says 'we'll learn them from what people already did.'

Cody The learned-versus-declared distinction is real. But I'll say it again — you need the behavioral data to be worth learning from. If your warehouse is chaos, query logs just give you a map of the chaos.

Justy Fair. But if you've got fifteen thousand contributors and three thousand production deployments on the open source side, there's a decent chance your deployment base isn't total chaos.

Cody Decent chance. Not guaranteed.

Justy Same! Shows you where the long tail actually lives. Alright Cody, go find your own golden queries today. Preferably before your own SQL starts hallucinating.

Cody That's… that's actually what I'm trying to avoid.