Agami Founders - Agami Insights

text to sql accuracy: 91.2% on Spider 1.0 with declared foreign keys against 21.3% on Spider 2.0 built from real enterprise warehouses, the same code agent on both

Data Agent

Why Text-to-SQL Accuracy Slips from 91.2% to 21.3% on Real Company Data

Text-to-SQL scores 91.2% on the benchmark everyone quotes and 21.3% on the one built from enterprise warehouses. The gap is implicit joins, and a semantic model is where they get declared.

jd edwards f0902 ledger type: the same twelve F0902 period columns hold dollars on the AA row, budget dollars on BA, and labor hours on AU, told apart only by the GBLT code

Data Agent

How to Sum JD Edwards Actuals Without Adding the Labor Hours to the Dollars

JD Edwards posts dollars, budget dollars, and labor hours into the same F0902 period columns, told apart by a two-letter ledger type. The semantic model names the ledger once, for every query.

workday hcm data warehouse headcount pre-hire worker same table: PERSONAL_INFORMATION, PERSON_NAME and ADDRESS each carry a column called type, and the rows it separates

Data Agent

In Your Workday HCM Warehouse, PERSON_NAME Holds People Who Don't Work for You

Fivetran lands Workday pre-hires beside workers and relatives beside employees, told apart by one column called type. The semantic model declares which rows each question means.

banner sgbstdn effective term data warehouse reporting: 109 columns and 99 correct foreign keys, and no column recording when a student record stopped being true

Data Agent

Ellucian Banner Records When a Major Started. No Column Records When It Stopped.

A replicated Banner lands every column and 99 correct foreign keys on the student record, and still cannot say how many students are in each major. The semantic model chooses the in-force row.

servicetitan job splits data warehouse: every foreign key correct and the technician revenue total still too big

Data Agent

The ServiceTitan Column That Stops One Ticket Paying Two Technicians in Full

Every foreign key in a replicated ServiceTitan schema is correct, and the obvious join still credits each technician the whole ticket. The semantic model is where the split weight gets applied.

restaurant business date vs calendar date sales reporting data warehouse: nine declared timestamps split the trading night, the tenth does not.

Data Agent

Ten Date Columns on a Toast Order, and Only One Knows When Friday Ended

Toast stamps every order with the business day its own reports use, then leaves the rule that computed it on an endpoint the connector never calls. The semantic model makes it the only day grain.

shopify presentment currency data warehouse: one money field lands as four columns, and only the shop-currency amount is safe to sum.

Data Agent

How to Sum Shopify Revenue When Every Money Column Lands Twice

Every Shopify money field lands twice, once in your store's currency and once in whatever currency each customer paid in. The semantic model binds revenue to one and refuses to sum the other.

query quickbooks custom fields fivetran warehouse: twelve configured on the invoice form, three returned by the REST API, and no column marking the nine that never arrived.

Data Agent

You Configured Twelve QuickBooks Custom Fields. Three Arrive.

QuickBooks Online Advanced lets you configure twelve custom fields per form. The REST API returns three. The semantic model records which three, and fails loudly on the nine that never arrive.

HubSpot duplicate deals multiple companies: deal_company is keyed on deal_id, company_id, type_id and category, so one deal on one company already lands two rows, each with the full amount

Data Agent

One HubSpot Deal, Three Companies, Three Rows

A HubSpot deal has no company column. The link lives in a bridge table that repeats per association type, so pipeline by account double counts. The semantic model declares the primary association.

Query Dynamics 365 Sales data with AI: the Won label the app shows and the integer 1 the warehouse lands. Every code field is an integer; labels live in stringmap.

Data Agent

Three Columns Mean Won. Your Semantic Model Picks One.

Every code field on the opportunity lands as an integer, and the labels live in a stringmap the connector cannot decode. The semantic model is where Won stops meaning three different things.

Query Acumatica custom fields Fivetran warehouse: the Priority Tier Attribute the app shows never lands. Custom data is unchecked on the connector.

Data Agent

Fivetran Lands 19 Acumatica Tables and Not One Attribute

Fivetran's Acumatica connector lands 19 tables and no Attribute values, so the fields the business asks about are not in the warehouse copy. What a semantic model can recover, and what it cannot.

Query Procore custom fields from your warehouse: what the app shows as a Risk Impact dropdown lands in the warehouse as a JSON blob whose key is an integer the reader has never seen.

Data Agent

The Procore Custom Field Whose Column Name Is a Number

Procore hands you custom_fields as a JSON blob keyed by integers. The label your team picked is nowhere in the warehouse copy. The semantic model is where that id gets bound back to its meaning.