Asking across systems
Ask one question that needs several systems at once, such as "which accounts with open renewals have overdue SAP invoices?", and get an exact answer: every row is counted, nothing is estimated, and you can see where each number came from.
Status: New. Tell support@sourcelace.com what you ask with it and how the answers look.
What it does
When a question needs more than one system, or more rows than fit in a chat, the AI writes the question once. SourceLace then:
- Asks each system only for what the question needs. Each system gets its own filters and fields (for example only open opportunities, or only invoices past due), with your own sign-in, so it returns only rows you are allowed to see.
- Asks every system at the same time. Three systems take about as long as the slowest one, not the three one after another.
- Puts the rows together and works out the answer exactly. The rows are joined, counted and totalled in a short-lived, encrypted place on SourceLace's server that only your question can read. The AI quotes the numbers it computes; it does not estimate them.
- Deletes the rows when the result expires, at the same time as any other result (30 minutes unless your admin chose less), or as soon as you close the chat if your admin set that.
A question to one system works the same way when its answer has more rows than a chat result holds (2,000): SourceLace keeps every row on its own, up to 250,000 (the number fetched without asking first), so totals and counts cover all of them. Past that, the result is cut, and the answer says it covers only the first rows; a question across systems can go further after you confirm the plan.
What you see
Under the answer's table, one line says where it came from, for example "Read from salesforce:dev, snowflake:snowflake-benchmark · all rows read · 10 rows". How it was read, at the end of the line, lists each system, the object read, how many rows came back and how long it took. Admins also see exactly what SourceLace sent to each system.
When the assistant read several tables to get there, the chat shows the tables you asked for and folds the rest under Built from; see Tables in answers, and Where this came from for where each column came from.
The line under the answer says all rows read when every row was read from every system, so totals and counts cover all of the data (it says nothing about whether the question was answered, only about how much data was read). If a system had more rows than your organization allows for one question, SourceLace stops and tells you how many each system has, so you can narrow the question. Only when you ask for an answer from the first rows anyway is it marked partial data, and the AI says so. The small ⓘ next to the label explains it.
You can keep asking about the same answer ("now by region", "which of these are in Germany?") without SourceLace fetching the rows again.
When the same customer or supplier is written differently in two systems, such as "ACME Corp." and "Acme Corporation GmbH", records are joined usually only on values that are exactly equal, unless your organization has a matching rule for the two objects. See Matching records across systems.
Large questions ask first
If a question would fetch a lot of rows (more than 250,000 unless your organization set another number), SourceLace shows the plan before anything is fetched: each system, about how many rows it would return, and how many of your organization's engine rows are left this month. In the SourceLace app, click Run this plan to go ahead; in your AI app, tell it to go ahead. Or narrow the question instead. A plan can be run for 10 minutes.
What protects your data
- Your own access. Every system is read with your own sign-in, and your organization's groups decide which objects you may ask about. They are checked again each time you ask more about an answer.
- Data protection first. Fields your organization masks, and personal data found in text, stay masked: the real values never reach the place where the answer is worked out. When a question cannot be answered exactly without the real values, SourceLace says so instead of giving a wrong number. An answer made in one chat can only be used in that chat: to use it in another, ask again there.
- Not kept. The rows are encrypted with a key of their own (or, on a server without encryption set up, held only in its memory, never on disk) and deleted when the result expires. Nothing is copied into a database. Every answer is read live from your systems, unless your organization has chosen Saved answers, an optional add-on your admin turns on system by system. Each system's part of the answer may then be one of your saved answers, and the answer says which.
- Audited. Your organization's audit trail records the question and, for each system, what was sent and how many rows came back, never the rows themselves.
Totals worked out by your system
A system connected through a REST or SOAP description, or from a vendor's own tools, can be one of the systems in the question. An admin can tell SourceLace which of its parameters narrows which field, so a long list is cut at the system (Filters you declare).
What it is, in one example
Suppose your database holds a million sales rows, and a region column that people typed in slightly different ways. Four of the spellings, with the total sold under each:
| Region as typed | Sold |
|---|---|
West |
100 |
WEST |
200 |
west |
400 |
west (a space at the end, which you cannot see) |
800 |
You ask: "total sales by region."
- The slow way: SourceLace fetches all million rows and adds them up itself.
- The fast way (what happens when your system supports it): SourceLace asks your database, "add these up for me, one total per region". The database sends back a few rows, not a million, and SourceLace finishes the answer from them. Your system does the heavy adding, and SourceLace gives the final answer.
The answer is the same either way, and the fast way is quicker, puts less load on your system and uses fewer of your row limits (see Limits). The answer's Sources show the request each system ran, with the note "aggregated at the source", so you can always see what was asked. Anything that cannot be done this way, and every other kind of system, is fetched as rows, as before.
Nothing to set up. It happens by itself and needs no change in your database. If you ever want it off for one system, an admin can do that (see "Switch it off for one connection" below).
Which regions are "the same"? SourceLace is strict; databases often are not
Every database has its own rule for when two pieces of text count as the same. For the four spellings above, a database used on its own would typically give:
| System, as usually set up | Groups it makes | Total for each |
|---|---|---|
| Microsoft SQL Server | 1: capitals and the space at the end are ignored | 1,500 |
| MySQL | 2: capitals are ignored, the space at the end is not | 700 and 800 |
| Microsoft Fabric | 3: capitals are kept apart, the space at the end is ignored | 100, 200 and 1,200 |
| Amazon Redshift | 3: the same as Fabric | 100, 200 and 1,200 |
| SourceLace | 4: everything is kept apart | 100, 200, 400 and 800 |
(These are the usual settings, each seen on made-up test data: MySQL 8.4 and SQL Server 2022 as they start, and Fabric and Redshift on real accounts. Both MySQL and SQL Server also treat "Zoë" written as one character and as "e" plus an accent as the same name; SourceLace does not. Each rule can be changed on a database or a single column, so yours may differ.)
SourceLace keeps all four apart on purpose, so that an answer never depends on how it was worked out. When it asks a database to add things up, it tells the database to compare text letter for letter for that one request, so the database gives the same groups SourceLace would. You get four groups, 100, 200, 400 and 800, whether SourceLace worked them out or the database did.
What you may notice. If you compare a SourceLace answer with a report run in your database's own query tool, SourceLace may show more groups. It is not an error: your database's tool merged spellings that SourceLace keeps apart. If you want them counted together, say so in the question, for example "total sales by region, treating West, WEST and west as the same region". That applies to that question only.
One exception. A database or column your administrators set up to ignore capitals (a "case-insensitive" setting; a Redshift database can be created that way, and a BigQuery, Snowflake, Databricks or PostgreSQL column can carry such a rule of its own) merges "West" and "WEST" itself. On PostgreSQL connections the rule of each text column is read from the database, and a column with a case-insensitive rule is then grouped letter for letter by the engine, with no setting needed; on the other systems that rule is not visible yet. If a total looks merged for that reason, an admin can switch the feature off for that connection (below).
Switch it off for one connection
An admin can stop a connection adding things up itself where MySQL, SQL Server, Fabric or Redshift compare text more loosely than the strict way described above, or where Snowflake, BigQuery, Databricks or a PostgreSQL database hold columns that ignore capitals on their own. It affects that connection only: other connections, and other organizations, are not touched.
- Open Connections and choose the connection. (For a database connection, the setting shows for MySQL, SQL Server and PostgreSQL.)
- Find Let this system work out totals and choose Off: SourceLace reads the rows and adds them up. Hover the ⓘ beside it for an example.
- Save.
To check it, ask for a total by a field of that system and open Sources under the answer. With the setting on, you see the note "aggregated at the source" and a grouped request. With it off, you see a plain read of the rows and no such note. The numbers are the same, except where a database or column is set up to ignore capitals, which is the case the setting is for; off is slower on big tables, because every row travels. Use it only when a total looks merged or the system struggles with the grouped request.
Compare text the way your database does
By default SourceLace keeps "West", "WEST" and "west" apart (see above). If your team's own reports come from a MySQL or SQL Server database that treats them as one region, an admin can tell SourceLace to follow the database's rule for that connection, so the totals match the reports your team already reads. This is a choice for one connection; nothing changes until an admin makes it.
Example. The same four spellings as above, in a MySQL database used as it is usually set up:
| Setting | Groups in "total sales by region" | Total for each |
|---|---|---|
| Off (the default) | 4: West, WEST, west and "west " | 100, 200, 400 and 800 |
| On | 2: West, WEST and west together; "west " apart | 700 and 800 |
In SQL Server, the same setting gives 1 group of 1,500, because it also ignores the space at the end. The group is shown under its smallest spelling in alphabetical order of the letters' codes (here "WEST"), so the same question always shows the same name.
What it covers. Totals by a text field, and the conditions "is", "is one of" and "is not one of" on a text field of one table. A question gives the same answer whichever way it was worked out, with the connection's totals switch on or off. What it does not change: joins between two systems (a "West" in your database still does not join a "west" in another system unless a matching rule links them), ranges, "starts with", sorting, counting different values, and any field you may only see masked. A question that groups a text field inside a WITH clause or a subquery and then filters on it is answered letter for letter, and the answer says so.
It follows the rule of each column. SourceLace reads the rule of every text column from the database. A column whose rule is one of the usual rules of these databases (such as one that ignores capitals) is compared that way; a column with a rule it does not know, such as a language-specific one, or one that already compares letter for letter, stays letter for letter. The answer's Sources says when a rule was used, such as "text in Sales DB is compared the way that system does: it ignores capitals and accents". Where a database also treats some letters as the same, such as "ß" and "ss", "ø" and "o", or Cyrillic "е" and "ё", SourceLace keeps them apart; the answer then has more groups than the database's own report, never fewer.
- Open Connections and choose a MySQL or SQL Server database connection.
- Find Compare text the way this database does and choose On. Hover the ⓘ beside it for an example.
- Save, then ask for a total by a text field and open Sources under the answer to see the note.
It does not make questions faster. Use it only when you want totals that match the database's own reports.
Which systems do this
SourceLace can ask a supported SQL database or data warehouse to work out totals, counts and the smallest or largest value of one table:
| System | Counts, smallest and largest values | Totals (sums) | Grouping by a text field |
|---|---|---|---|
| PostgreSQL, BigQuery, Athena, lake tables | available | available | yes |
| Snowflake, Trino | available | whole numbers only | yes |
| Databricks | available | not yet | yes |
| Amazon Redshift | available | available | yes, for varchar columns (see the exception above) |
| MySQL | available (not for dates) | available | yes, for varchar and text columns |
| SQL Server | available | available | yes, for nvarchar and nchar columns |
| Microsoft Fabric | available | available | yes, for varchar columns; not for Azure Synapse pools |
| Oracle | available | available | no |
Which systems were checked on a real system: Google BigQuery, MySQL (on Azure), Microsoft Fabric (a warehouse) and Amazon Redshift (Serverless) were each checked with made-up rows on a real system before this was switched on; PostgreSQL and SQL Server are tested on real servers every time SourceLace changes. A Fabric lakehouse's SQL endpoint shares an address with a warehouse and was not checked separately.
Which systems can be combined
In a question across systems you write each table as "source"."object". Which systems can be named that way:
| Can be named as a table | Takes part in two steps | Cannot be part of a question across systems |
|---|---|---|
| Salesforce, HubSpot, Dynamics 365, NetSuite, Oracle Fusion, Workday, SAP S/4HANA, SAP ECC, SAP SuccessFactors, SAP Datasphere, Data 360, ServiceNow, Jira, Gainsight, ChurnZero, Dynamics 365 Finance and Supply Chain, Flexport, claims (a remittance advice, named by its id), MongoDB, SQL databases, Snowflake, BigQuery, Databricks, Amazon Redshift, Athena, Microsoft Fabric, Trino, lake tables, Salesforce Marketing Cloud (a data extension, named by its key), Google Sheets (a tab, named as <spreadsheet id>/<tab name>), FHIR health records (a resource type such as Observation) and Custom connectors that offer table reads (see Custom connectors); REST, SOAP and outside tool (MCP) sources, for an operation or tool that needs no value from you | Gong, Zendesk, and Custom connectors that do not offer table reads | Files, mail, chat and search (they hold files or messages, not tables) |
Two steps means this today: ask the system on its own first, with a normal question (SourceLace holds the answer as a result), then name that result in the question across systems as "results"."<result id>". The assistant does this for you; when you write the SQL yourself in an AI app, follow the message in the table below. The result is joined exactly like any other table.
When something goes wrong
| What you see | What it means | What to do |
|---|---|---|
| "<table> cannot be read in a query across sources yet (<system>). Query it on its own with query, then name its result in this query as "results"."<result id>"." | The system is one that takes part in two steps (Gong, Zendesk, or a Custom connector that does not offer table reads). Its table cannot be named directly. | Ask that system on its own first, then name its result in the question across systems, as the message shows. |
| "<source> (<label>) holds files, mail or messages, not tables, so it cannot be part of a query across sources. Use query on it instead." | Files, mail, chat and search hold documents or messages, not rows. | Find or read them with a normal question, then ask about what you found. |
| "<table> has no fields SourceLace can read. Check it with describe_object." | The table or operation is known, but SourceLace could not see any field in it. | Ask the assistant to describe the object to see what the system reports; if it reports no fields, ask your admin to check the connection, then ask again. |
Limits
How many systems one question may use, how many rows each may return, how long a question may take and how many rows your organization may fetch each month are on the Limits page.