SourceLace Docs
FAQOpen the app

Limits

How far SourceLace goes for your organization: how much work the AI may do for one question, how long agents may run, how long people stay signed in, how long queries may take and how large a file SourceLace reads. Admins see every limit on the Limits page, grouped under headings such as AI answers, AI agents and Across systems, and can make some of them stricter. You can search all of them, show only the ones you can change, and see which you have changed.

Who sets a limit

Your plan sets each limit, and your SourceLace contact can set a different value for your organization (for example for a pilot). That value is what Your plan allows on the Limits page.

For the limits with a box in the Your value column, an admin can choose a stricter (lower) value, for example shorter sign-in sessions for your security policy, or fewer agent runs at once. You can never choose more than SourceLace allows; ask your SourceLace contact for more. Leave a box empty to use what SourceLace allows. A box that holds a value outside the allowed range is marked, and Save limits waits until it is fixed; Undo puts one box back and Discard all of them. Changes apply within a minute, and a shorter sign-in session also ends sessions that are already older than it.

Some limits are set by SourceLace only, such as the longest answer the AI may write. The page shows them so you know them.

Finding your way on the page. Ten topics are listed on the left (a drop-down on a phone), each with the number of limits in it. Choose one to see its limits. Type in the search box to look across all of them, for example "agent", "minute" or "files"; ticking Only ones I can change or Only ones I changed narrows the list the same way. Each row shows the limit, Your value, Your plan allows and its state: SourceLace's value, Stricter (yours) (with Use SourceLace's to go back), Unsaved or Set by SourceLace. More shows the full description and the allowed range. Times and sizes are shown in minutes, days, MB or GB. A bar at the bottom counts the changes you have not saved yet and holds Save limits and Discard. Every save is recorded in the audit trail. Members who open the page see only a message that it is for admins.

Two more limits come with your plan or pilot but are not on the Limits page:

  • Changes to records. A plan without them refuses every change (SourceLace only reads), whatever an admin has allowed on Connections. A pilot includes them, and each change still needs approval.
  • Sources in all. A plan or pilot can cap the total number of sources, whatever their connector class (a pilot allows 5 by default). At that number, adding a source is refused until one is removed.

When a system is busy

SourceLace keeps a request budget for each of your systems (requests per minute and at the same time, set for each kind of system, such as Salesforce's API limits) and for the AI, so that a busy question, person or agent never floods a system and your other tools keep their share. The budget is shared fairly: while someone is waiting, nobody may take more than their share of it, so one busy person or agent cannot hold up everyone else. With SourceLace's included AI, each organization also gets its fair share when many use it at once.

When a budget is full, a request waits a few seconds for room. If there is still none, it is answered with a message such as "Salesforce is busy: SourceLace is already sending it all the requests it allows right now. Try again in 20 seconds." or "The AI model is busy right now. Try again in 15 seconds." Nothing is lost: try again after that many seconds. Your AI app can do so by itself. An AI agent that finds a system or the AI busy before it has changed or drafted anything goes back in the queue and starts again later on its own; see When a run fails. When a system itself asks SourceLace to slow down, SourceLace waits as long as it asks.

Admins see on the Usage page on which days your organization was slowed this way. If it happens often, spread out agents' schedules, or ask your SourceLace contact about your limits. The limits below that decide it are Model calls per minute, Model input tokens per minute, Model output tokens per minute, Share of each system's request budget, Wait when a system is busy and Tries when a system is busy.

Every limit

Limits are listed in the order they appear on the Limits page, under the same headings.

AI answers

How hard the AI may work on one chat question, and how much your people may use per minute.

Model calls per question

How many times the AI may call the model to answer one chat question (looking up schemas, running queries, joining). More answers harder questions and costs more. An admin can choose a lower value.

Lookups at the same time per question

When the AI needs several things at once to answer a question, such as the fields of three objects or a query in two systems, it looks them up at the same time instead of one after another, so answers come sooner. This is how many lookups run at once. Changes and documents always wait until the lookups are done and then happen one at a time, in order. If one lookup fails, the others still finish. An admin can choose a lower value; 1 does every lookup in turn.

Output tokens per model call

The longest answer one model call may write.

Model calls per minute

The most calls your organization's AI makes to the model in a minute: chat, agents and setup together, shared fairly between the people asking. With your own AI key it is that key's budget for each model, as AI providers count each model separately, so set it to what your provider account allows. A call over it waits a few seconds, then is answered "busy, try again" (see When a system is busy). An admin can choose a lower value.

Model input tokens per minute

The most tokens your organization's AI sends the model to read in a minute, for everyone together, shared fairly between the people asking. As with AI providers, what the model reads again from its cache (the instructions and the earlier turns of a conversation) does not count, so long conversations use little of it. An admin can choose a lower value.

Model output tokens per minute

The most tokens your organization's AI writes in a minute, for everyone together, shared fairly between the people asking. An admin can choose a lower value.

AI agents

What one agent run may do, how agents are started and shared, and when a change needs a person.

Steps per agent run

The most tool calls one agent run may make. Each agent's own limit must be at most this. An admin can choose a lower value.

Credits per agent run

The most credits (AI use plus steps) one agent run may use. Each agent's own limit must be at most this. An admin can choose a lower value.

Time per agent run

How long one agent run may take before it is stopped. An admin can choose a lower value.

Chat messages per agent run

How many Slack or Teams messages one agent run may post or propose. An admin can choose a lower value.

Agent lookups at the same time

Like Lookups at the same time per question, for agents: how many of one run's lookups run at once, so runs finish sooner. Record changes, email drafts and chat messages always wait until the lookups are done and then happen one at a time, in order. Each lookup still counts as one step, and a run never makes more steps than its limit. An admin can choose a lower value.

Tries when a system is busy

When an agent run finds a system or the AI busy before it has changed or drafted anything, it goes back in the queue and starts again later, up to this many times (5 by default), waiting longer each time. 0 ends such runs as failed at once. An admin can choose a lower value.

Agent runs at the same time

How many of the organization's agent runs may run at once, across all servers, so one organization cannot take every place. Others wait for the next turn. An admin can choose a lower value.

Agent deliveries at the same time

How many of the organization's agent deliveries (files attached to email drafts or posted in chats, and changes tried again after a system asked to wait) may be sent at once, across all servers, so one organization cannot take every place. Others wait for the next turn. An admin can choose a lower value.

Webhook calls per agent per minute

How many times other systems may call one agent's start address (its webhook trigger) in any minute; more are answered 429 (try again later). Calls that are refused or filtered out count too. An admin can choose a lower value.

Event runs per agent per minute

How many runs one agent may start in a minute from Salesforce events or from a message queue. More events or messages wait and start as room comes back; none is dropped.

Salesforce event streams per organization

How many streams of events your organization may have at once, from Salesforce events and from message queues: one for each Salesforce event topic and each person who listens to it, and one for each message queue topic that agents start from. Agents beyond it wait until a stream stops.

Emails per agent per minute

How many emails to one agent's email address may start it in any minute; more are dropped and shown in Recent emails. Emails that are refused or filtered out count too. An admin can choose a lower value.

A2A calls per person per minute

How many A2A calls (send a message, read, list or cancel a task) agents on other platforms may make in any minute signed in as one person, across the organization's agents; more are answered 429 (try again later). An admin can choose a lower value.

Subscribers per shared agent

How many people one shared agent may run for: people who subscribed, and the members of groups an admin subscribed. Each subscriber's run is metered like any agent run. An admin can choose a lower value.

Days a change waits for approval

How long a change an agent proposed waits for someone to approve it before it expires. An admin can choose a lower value. The agent's owner or an admin can extend one waiting change, never past 30 days from when it was proposed.

Time before the first approval reminder

How long after a change starts waiting for approval the people who must decide it are reminded by email, and from when the Approvals inbox marks it overdue. 0 switches the reminder off. It must not come after the warning point below. An admin can choose a lower value. See Approval emails.

Days before expiry to warn

How many days before a waiting change expires the people who must decide it are warned by email, and from when the Approvals inbox marks it expiring soon. 0 switches the warning off. It must be fewer than the days a change waits for approval. An admin can choose a lower value.

Reviewed shadow runs before going live

How many runs of an agent in shadow mode people must review before it may go live. Here more is stricter: an admin can choose a higher value, never a lower one.

Days in shadow before going live

How many days an agent must stay in shadow mode before it may go live, on top of its reviewed runs. 0 means no minimum. Here more is stricter: an admin can choose a higher value, never a lower one. Admins can also set this on Agent settings; the higher of the two counts. See Days in shadow and approved changes.

Most rows an agent may read without asking

The highest number an admin may give one agent as its Rows this agent may read without asking (2 million by default). Nobody can confirm a large question while an agent runs, so this lets an admin pre-approve agents whose job needs a large read. The rows per table, per question and per month still apply. An admin can choose a lower value. See Limits.

Days of agent performance history kept

How long the daily figures behind Agents → Performance (counts and durations, never your data) are kept: 400 days by default, so you can compare a year. The page itself looks back at most 90 days. An admin can choose a lower value. See Measuring your agents.

Runs before a success rate is shown

A success rate is shown only from this many finished runs (3 by default), so one run does not read as 0% or 100%. An admin can choose a lower value.

Runs before a hint about outcomes

The Where to improve hints about success, changes not made and failing or slow steps need this many finished runs or step calls in the period (10 by default). An admin can choose a lower value, to see hints sooner.

Approvals decided before a hint about approvals

The hints about approvals expiring or going stale need this many approvals decided in the period (5 by default). An admin can choose a lower value.

Hint thresholds

Each hint on Agents → Performance fires at a share you can tune: waiting for people at least this share of the journey (0.5), queue wait at least this share (0.3), success rate under this share (0.8; here an admin may raise it), runs ending without a change at least this share (0.3), approvals expired (0.2) or stale (0.1) at least this share of those decided, a step failing at least this share of its calls (0.2), deliveries failing at least this share (0.1), and a slow step whose slowest tenth takes longer than this many seconds (60). Lower values show hints sooner. The rules themselves are in Measuring your agents.

Looking up one system

How long and how much one lookup in a single system (database, warehouse, lake, web service) may take.

Query time

How long one query to a database or data warehouse (SQL databases, MongoDB, Snowflake, BigQuery) may run before SourceLace stops it. An admin can choose a lower value.

API requests per minute

The most requests SourceLace sends one REST or SOAP source in a minute, for everyone together, so a busy AI cannot flood the organization's system. An admin can choose a lower value.

Pages read per question (REST and SOAP)

The most pages of an answer SourceLace reads from one REST or SOAP source for one question. A longer list is cut there and the answer says it is incomplete. Every page also counts against the source's requests per minute, shared by everyone. An admin can choose a lower value.

Pages read per question (outside tool sources)

The most pages of an answer SourceLace reads from one outside tool source for one question. A longer list is cut there and the answer says it is incomplete. Each page is a call to the tool and counts against the source's requests per minute, shared by everyone. An admin can choose a lower value.

Fetches at once per source

How many fetches SourceLace runs at the same time against one source, for everyone in the organization together, so one busy question cannot flood a system. An admin can choose a lower value.

Share of each system's request budget

How much of the request budget SourceLace sets for each kind of system it uses for one of your sources: the requests per minute and at the same time it sends, for everyone in your organization together, shared fairly between the people and agents using it. 1 (the default) is SourceLace's budget for that kind of system, below the vendor's own limits; lower it to leave more room for your other tools that use the same system. See When a system is busy. An admin can choose a lower value.

Wait when a system is busy

How long a request waits for room when a system's budget or the AI's is full, before it is answered "busy, try again" with the seconds to wait: 5 seconds by default. 0 answers at once. An admin can choose a lower value.

Most bytes one lake query reads

How much one query on a lake source may read from storage, data and metadata together. A query that would read more is stopped. An admin can choose a lower value.

Most files one lake query reads

How many files (data and metadata) one query on a lake source may open. An admin can choose a lower value.

Health record results kept

How long SourceLace holds a result read from a health record (FHIR) source before deleting it: 5 minutes unless set otherwise, at most 30. These results hold protected health information, so they are kept for less time than other results; your data retention settings can shorten it further, never lengthen it. An admin can choose a lower value. See Health records.

Across systems

Questions that combine several systems: how many tables, rows and seconds one may use (Asking across systems). An admin can choose lower values for each of these except the monthly rows.

Tables in one cross-source query

How many tables (objects) one question across systems may read. 8 by default.

Rows fetched per source table

The most rows one table may return for a question (2 million by default) and the most from all tables together (5 million). A question that needs more is stopped with the count from each system, never cut short without saying so. When a supported system works out a total, count, smallest or largest value itself (Totals worked out by your system), the rows that actually travel count: the result, not the size of the table. Such a result is never cut short: when it would pass the limit, the question is stopped and says so.

Rows fetched per query

The most rows one query fetches into the engine from all its sources together, counting a total a system works out itself as the rows of its result. An admin can choose a lower value.

Rows fetched without asking first

Above this many rows (250,000 by default), SourceLace shows the plan and fetches only once you confirm. A total a system works out itself counts as the rows of its result, so such a question rarely needs confirming.

Largest held result

The most disk one held result may take in the engine (encrypted, deleted when it expires). An admin can choose a lower value.

Pairs one matching way may compare

Before a matching rule runs in a query across sources, SourceLace counts the pairs of records each way would compare (those that agree on its Same checks). Above this, the question is refused with the way and how to narrow it. An admin can choose a lower value.

Time per cross-source query

How long one question across systems may take, fetching and computing together (5 minutes by default). If it runs out, SourceLace says which system was still answering and how far each had got.

Time to estimate row counts

How long SourceLace asks sources for row counts before a large fetch; a source that does not answer in time shows as unknown. An admin can choose a lower value.

Time per analysis of a held result

How long one more query over a result SourceLace already holds may run. An admin can choose a lower value.

Sources fetched at once per query

How many of one question's tables SourceLace reads at the same time (6 by default), and how many reads run at once against one system for your whole organization (2 by default), shared fairly between your people, so one busy question cannot flood a system.

Cross-source queries at once

How many queries across sources the organization may run at the same time. An admin can choose a lower value.

Engine rows per month

How many rows your organization may fetch for questions across systems each month (engine rows), set by your plan. They use no credits. A total a system works out itself counts only the rows of its result. Admins get one email when 80% are used, and the Usage page shows how many are left. When none are left, questions that need to fetch more are refused until the 1st; to get more, ask SourceLace to raise your plan's engine rows. See Engine rows.

Files and folders

Reading files and searching inside them in Drive, SharePoint, S3, Azure and Google Cloud Storage.

Largest file read

The largest file SourceLace downloads to read (Google Drive, SharePoint, OneDrive, S3, Azure Storage, Google Cloud Storage, Slack, Teams). An admin can choose a lower value.

File names one search looks through

How many file names one search in Amazon S3, Azure Storage or Google Cloud Storage looks through: 10,000 unless set otherwise. A search that stops there says so and is marked incomplete; searching inside one folder looks through fewer names and is faster. An admin can choose a lower value. See Finding files faster.

Folders listed at once per source

How many folders SourceLace lists at the same time in one Amazon S3, Azure Storage or Google Cloud Storage source, for everyone in your organization together, when it searches file names. More finds files faster. An admin can choose a lower value.

Folders listed at once

How many folders SourceLace lists at the same time across all of your organization's Amazon S3, Azure Storage and Google Cloud Storage sources. An admin can choose a lower value.

Files one content search reads

How many files one search inside file contents reads, in an Amazon S3, Azure Storage or Google Cloud Storage source whose admin chose Names and file contents: 200 unless set otherwise. A search that would read more stops, answers with what it found, says so and is marked incomplete: choose a smaller folder. An admin can choose a lower value. See How S3, Azure Storage and Google Cloud Storage are searched.

Bytes one content search reads

How much one search inside file contents downloads in all: 50 MB unless set otherwise. A search that would read more stops the same way. An admin can choose a lower value.

How long one search inside file contents may read files: 20 seconds unless set otherwise. A search still reading then stops, answers with what it found and is marked incomplete. An admin can choose a lower value.

Search everything

How long and how many results the cross-system search asks each system for.

Search everything: time for each system

How long a search waits for each system's answer: 8 seconds by default. A system that takes longer is left out of that answer and named, and the answer says it is incomplete, so one slow system never holds up the others.

Search everything: results from each system

How many results a search asks each system for: 10 by default.

How many systems one search asks at the same time: 10 by default. When more of your systems can be searched, name the ones to search.

Sign-in and provisioning

How long people stay signed in, and how fast your identity provider may send updates.

Web app session

How long someone stays signed in to the SourceLace web app before signing in again. An admin can choose a lower value.

AI app sign-in (refresh token)

How long an AI app (MCP) stays signed in to SourceLace without the person signing in again. An admin can choose a lower value.

AI app access token

How long one access token for an AI app lasts before the app renews it. An admin can choose a lower value.

SCIM requests per minute

How many requests the organization's identity provider may send to SourceLace's SCIM endpoint in any minute; more are answered 429 (try again later), which identity providers retry.

Credits, budgets and warnings

What credits cost, and when admins are warned about use, seats, budgets and quiet runners.

Credits per agent step

Credits each tool call of an agent run costs, on top of its AI use.

Credits per question

How many credits make one "question" when credits are shown to people.

AI budgets

How many monthly AI budgets (for groups and agents) the organization may set.

Warn at this share of included AI use

Admins see a notice when the month's included AI use reaches this share (0.8 is 80%). An admin can choose a lower value.

Warn at this share of the extra maximum

Admins see a notice when extra AI use reaches this share of their monthly maximum. An admin can choose a lower value.

Warn at this share of seats

Admins see a notice when this share of the plan's maximum seats is used. An admin can choose a lower value.

Warn at this share of an AI budget

Admins (and an agent's owner) are emailed and warned when a group's or agent's monthly AI budget reaches this share (0.8 is 80%). An admin can choose a lower value.

Runner quiet alert after

How long a runner with agents assigned to it may go without asking SourceLace for work before the organization's admins see an alert. A runner asks every 15 seconds. An admin can choose a lower value.

Organization capacity

How many groups, managers, skills and on-premise agents the organization may have.

Access groups

How many access groups the organization may have.

Managers per group

How many people may manage one group: add and remove its people without being admins. See Group managers. An admin can choose a lower value.

Skills

How many skills the organization may have, and each person, and each project.

On-premise agents

How many SourceLace Agents (for on-premise systems such as SAP ECC) the organization may register.

Business setup

Business terms, matching rules, starter questions, and how the setup assistant and learning from use work.

Business terms

How many business terms (such as ARR or open pipeline) the organization may define.

Business terms per group

How many business terms one group may give its own meaning. See Group terms, starter questions and default sources. An admin can choose a lower value.

Matching rules

How many entity matching rules the organization may have (Business setup, Matching): one per pair of objects, drafts included. An admin can choose a lower value.

Starter questions

How many starter questions (the example questions in an empty chat) the organization may set for everyone. An admin can choose a lower value.

Starter questions per group

How many starter questions one group may set for its people. An admin can choose a lower value.

Objects listed per setup scan

How many objects (tables, entities) of one source the setup assistant lists in one scan.

Objects described per setup scan

How many objects of one source the setup assistant reads the fields of in one scan. An admin can choose a lower value.

Sample rows per object in a setup scan

How many rows the setup assistant reads per object to count values that look personal. They are counted in memory and never kept. 0 turns sampling off. An admin can choose a lower value.

Setup scans per day

How many setup assistant scans the organization may start in one day (UTC).

Days suggested setup changes are kept

How long suggested setup changes stay in the review queue before they are deleted.

People before a pattern is suggested

How many different people must ask similar questions before learning from use suggests a skill or agent for them. Higher is stricter.

Questions before a pattern is suggested

How many similar questions there must be before learning from use suggests a skill or agent.

Days of questions learning looks at

How far back learning from use looks for similar questions. Never further back than the organization keeps chats.