INSTALL.md).Query results never leave this machine. AI CaseQuery sends a question to your firm's own AI provider account, gets back a query, runs that query read-only against your database on the user's own computer, and composes the answer there.
What does go to your AI provider: the question, table and column names, and your firm's notes, terms and rules. There are two named exceptions to "no values":
Columns you hide under Schema rules are never sent. Chat history is kept only in each user's own browser. A provider key saved on a workstation is sealed to that computer or user and does not travel with a copied folder. A key saved once for a shared installation is kept in the folder in a form every workstation can open until you press Clear the shared copy, so until then a copy of the folder carries it too.
The Installation Guide (INSTALL.md) already covers installing Node.js, extracting the application, the first launch, the evaluation period and licensing, provisioning the AI key onto workstations in bulk, installing new versions, and installation troubleshooting. This guide does not repeat it. Start here once the application opens in a browser.
| Key | What it opens | Who holds it |
|---|---|---|
| Administrator key | Every tab under Settings, writing business rules and terms, confirming or withdrawing settled definitions, and the DBA screens. | Your IT team or nominated administrator. Do not circulate it to users. |
| Leads' key | The lead tier on Fine-tuning: settling definitions, marking baseline answers, handing questions to the DBA. It does not open Settings. | Operations Leads. Optional: a firm with no separate lead tier can leave it out. |
Instead of keys, a firm can grant both tiers through its own Windows groups or through Microsoft Entra ID, using the roles AI CaseQuery Administrator and AI CaseQuery Operations Lead. You choose this under Settings › Security › Who may administer (see 3.3).
An unlock lasts until you close that browser tab. Click the badge to lock again at any time, for example before someone else uses your screen. If the application restarts, the unlock ends and Settings asks again.
Leads unlock from Fine-tuning with Unlock to settle definitions, or are prompted the first time they act.
Settings › Setup checklist lists what to do, in order. The heading reminds you to lock down the program folder and the settings files last. Several settings are written to the data folder so that every workstation reads the same value, and locking it early means unlocking it again to finish.
Each step shows one of two labels on the right:
Open beside a step title takes you to the screen where it is done. The progress line counts steps done. Optional steps, and the business-rules step, do not hold the checklist open.
Some steps only appear in certain installations. They are marked below.
Why: the folder's permissions are the strongest control on who can run the application at all. The step shows where the installation is: the program in program\, your settings and data in data\.
How: to stop anyone else running it, restrict Read on the installation folder to a group containing only the people who should use it. Removing only Execute achieves nothing, because Read is enough to copy the folder and run it elsewhere. Then confirm.
Why: until someone approves the application for your organisation, nobody can use it.
How: the first administrator to sign in with a work account is asked to approve it. The step ticks itself once someone has signed in. Open goes to Security.
Why: someone who signs in successfully but holds no role can configure nothing, and the screen cannot explain that the fix is in your own admin centre.
How: in the Microsoft admin centre, open Enterprise applications, find this application, and assign people to AI CaseQuery Administrator or AI CaseQuery Operations Lead. Assigning a group rather than individuals needs Microsoft Entra ID P1. The step ticks once the signed-in account holds a role.
Why: shown only during an evaluation where nothing else has been arranged. The evaluation keys stop working the day a licence is applied.
How: set your own access key, Windows group or Microsoft Entra ID under Security › Who may administer, and test it while the evaluation keys still work. Then confirm.
Why: removing someone from the group then removes their access through the process you already use when people leave.
How: add the right people to the mapped groups in Active Directory or on the machine. Windows reads group membership at sign-in, so someone just added must sign out and back in. The step ticks when your own account holds a role through a group.
Why: only useful if your firm collects the Windows Event Log, as most firms with a security monitoring system do. The activity journal is written either way.
How: the step shows a command to run in Windows PowerShell started with "Run as administrator", on each workstation. It needs elevation, which the application deliberately does not have. Press Check again afterwards. Once registered, sign-ins, refusals, exports and configuration changes go to the Application log as well. Confirm the step if you have run it or decided not to.
Why: the sample connection, the sample's rules and notes, and foreclosure-style starter questions would otherwise stay one click away in front of your staff or clients.
How: close the application and run Reset.bat in the program folder. It removes the sample connection, the rules and notes that came with the sample database, and the evaluation history. It keeps your licence, provider keys, access keys, the audit journal and the sample database itself.
Why: without publishing, every user has to configure the database themselves.
How: Open goes to Database. Choose the Database type, fill in the connection, press Connect read-only, then under Make this connection live for everyone press Publish this connection (see 3.7). The step ticks when a connection is both connected and published.
Why: a hidden column is never read when the database is prepared, so nothing about it is ever written down. Hide it afterwards and a note describing it may already exist on disk.
How: Open goes to Schema rules. Review what is flagged, tick only columns nobody should see, and press Save & lock policy (see 3.9). Hiding nothing is a legitimate choice, because a hidden column also stops the AI using it. Save the policy even with nothing ticked, so the policy file exists for step 16. Then confirm.
Why: preparation describes the coded columns so the AI does not guess at them, and lists the definitions your firm needs to settle. Users cannot ask questions until it has run.
How: Open goes to DB Tuning › Prepare the database (see 2.2). The step ticks when preparation has run, and shows how many column notes were written and how many definitions there are to settle. If the database's structure changes later, the step opens again and says what changed.
Why: if the application already answers for you, that key may be in your own browser and apply to you alone. Every other user would have to paste their own.
How: the key must not stay in the application folder, which is designed to be copied. It is sealed by Windows onto each authorised computer instead.
For many workstations, the Installation Guide describes unattended provisioning from Group Policy or Intune. An environment variable also works (see 3.3). The step ticks when a key is in force from a source other than a personal browser key.
Why: leaving the defaults is a real decision, and it looks identical to never having opened the screen. The step shows the current row limit, AI response limit, query limit and whether question checking is on.
How: Open goes to Safeguards (see 3.4). The defaults are sensible. Confirm when you are happy with them.
Why: the AI provider, model, thinking level, fast mode and safeguards you set on your own PC apply to you alone. Until you make them the installation's starting settings, a new PC or Windows user with no choices of their own starts on the settings the application ships with, not yours. This is the step people forget, which is why the warning bars and this step both point to it.
How: Open goes to Security › Who may administer › Installation settings (see 3.3). Press Make these the installation's starting settings. Do it before copying this folder to other PCs: each copy keeps the settings it had when it was copied. The step ticks once they are set; if your own settings later differ, the step says where.
Why: the assistant, Casey, explains the product and uses your AI provider account. Whether Casey may also read what your firm has written down is a separate permission people miss.
How: Open goes to Assistant (see 3.6). Decide both switches, then confirm.
Why: rules are things only your practice knows, such as what counts as referred. They come from the teams using the product, not from whoever installs it. Invented rules are worse than none: they steer every future answer and nobody remembers agreeing to them.
How: nothing to do now. Tick it and leave it. Rules arrive later when someone meets a number that looks wrong and reports it from the answer. Open goes to Business Rules.
Why: anyone who can change data\deployment.json can make themselves an administrator, however the keys are stored. data\security-policy.json decides which columns are hidden. Changing a file in program\ is changing the application.
How: do this only when everything above is done.
program\.data\deployment.json and data\security-policy.json. Both files must exist first: saving any setting creates the first, saving the hidden-column policy creates the second.data\ writable, with Modify, not Full control. The application writes rules, settled definitions, column notes and preparation results there as people use it. Full control on the folder would let someone delete a protected file and put their own in its place.The step reads the permissions and names anything standard users can still change. It sees the common Windows groups (Everyone, Users, Authenticated Users), not every named account or custom group, and not share permissions on a network folder. Control those as well.
Reset.bat, applying a licence, changing hidden columns, publishing a connection, changing the corrections folder or changing who may administer all write to the protected files. Do them while signed in to Windows as an account allowed to change those files.| Guide | Give it to |
|---|---|
| Admin / DBA Guide (this guide) | Administrators, and the DBA who works with them |
| Lead Guide: Fine-tuning | Operations Leads |
| User Guide | Everyone who asks questions |
| Working with Casey | Everyone, if Casey is offered |
Six kinds of tuning shape answers. They are different tools with different reach.
| What | Where | Who writes it | What it changes in an answer |
|---|---|---|---|
| Column notes | Schema rules; written first by preparation | Administrator | How the AI reads an abbreviated or coded column. Sent with questions as part of the schema. |
| Business rules | Business Rules › Rules | Administrator | A standing instruction ("always do Z"). Applies to every question it concerns, including ones nobody has asked yet. |
| Terms | Business Rules › Terms | Administrator | What a phrase means ("when someone says X, they mean Y"), optionally with the exact condition and tables. |
| Settled definitions | Fine-tuning › Definitions to settle | Lead settles; administrator confirms or withdraws | Which reading of an ambiguous word or measurement the firm means. Applies from the moment it is settled, exactly as a rule does. |
| Baseline questions | Fine-tuning › Baseline questions | Lead marks answers | Nothing directly. A record of which answers were right on a given date, and a source of work: a Wrong answer becomes a definition to settle, and a Right answer is offered to the DBA as a pair to settle. |
| Verified queries (Question-SQL pairs) | DB Tuning › Question-SQL pairs | Administrator, with the DBA | That exact question is answered by running your query, word for word: no AI call, no wait, and the same query every time (the figure follows your data as it changes). |
| Ground truth | DB Tuning › Ground truth | Administrator | Nothing directly. A check of answers against figures your firm already trusts, run when you choose. |
Up to a dozen rules, terms and settled definitions together are sent with every question. Past that, each travels with the questions it concerns: a rule naming a table goes with questions about that table, and a rule naming nothing is always sent. Tick Always apply on a rule to send it with every question whatever it names — for a firm-wide rule that happens to mention one table.
Go to Settings › DB Tuning › Prepare the database. Preparation reads the connected database once. It describes the coded columns so the assistant is not guessing, and finds the definitions your firm needs to settle. It runs locally; table data is not sent to the AI. The profile it keeps holds structure, counts and characteristics, not your records.
.sql, Word .docx or Excel .xlsx). Press Read it. Only the documentation is sent through your AI account, never database records. The report shows where your documentation and your database disagree, which statements the data confirmed, which were taken from your documentation without a check, and which could not be checked. The report is for you to act on; nothing from it is saved.Press Prepare Database. Preparation times each query and eases off while the database is busy, so it is safe on a live system and quicker outside peak hours. Stop ends a run without confirmation, because nothing is written until preparation is applied. When it finishes, read the report before anything is written.
The badge beside the heading reads Not run, Prepared, Skipped or Schema changed — run again.
Leads settle or set aside definitions, mark baseline answers, run baseline sections, add baseline questions, use Check the wording and See what this changes, and hand questions to the DBA with Send to the DBA. They cannot open Settings, save rules or terms, or confirm or withdraw settled definitions. Their day-to-day work is described in the Lead Guide: Fine-tuning.
Every recorded decision asks for a name in a Your name box before it is saved.
Go to Fine-tuning › 3 · Awaiting sign-off. It lists the definitions a lead has settled that are already in force and still waiting for your second signature. Confirmed definitions are on Business Rules › Rules › Definitions in effect from Fine-tuning, where a confirmed one can be withdrawn; withdrawn and cleared ones are on Business Rules › History.
| State | Meaning |
|---|---|
| provisional | Settled, not yet reviewed. It applies. |
| stale | Unreviewed for over 30 days. It still applies. |
| archived | Unreviewed for over 90 days. It still applies. |
| active | Reviewed and confirmed by an administrator. |
| withdrawn | No longer applied. |
While you are unlocked as administrator, the unreviewed ones are pinned at the top under a heading saying how many definitions the firm has not stood behind. For each:
Both ask for your name every time and never reuse a remembered one, because confirming is a second signature by someone other than the lead who settled it. The same definitions also appear on Business Rules › Rules › Definitions in effect from Fine-tuning, marked In force · awaiting sign-off, because they already apply. A header badge appears on every screen while reviews are owed.
Business Rules › Has it worked? is observed from questions people actually asked. It is advisory only: nothing here changes an answer.
A question settling shows the wording stopped being ambiguous. It does not prove which rule did it.
Floor voting asks the people doing the work what an ambiguous word means, instead of the product choosing quietly. It is off by default.
Go to Settings › Safeguards › On Floor Voting and switch on Ask the floor when a question is ambiguous.
A vote appears under an answer only when both are true:
It reads, for example, Counted "open" as Open. Is that right?, with a button per value and Not sure. A line says who will see the vote.
Nothing is asked about "anything except", "one of" or "like" filters, numbers or yes/no flags, or queries filtering more than one status column. A column with 7 or more values, or a person who has switched asking off, gets only the line Counted "active" as Active. with nothing to vote on. Questions with a Verified query never ask.
| Setting | Default | What it does |
|---|---|---|
| Votes needed to settle it | 4 (2 to 50) | When that many different identified people agree, it becomes a provisional settled definition that applies straight away. You then confirm or withdraw it under Awaiting sign-off. A small firm may only have three or four people who would know. |
| Stop asking after | 10 (2 to 200) | Once this many people have answered without agreement, nobody else is asked and it goes to the leads. Only answers count; being shown the question does not. |
After voting, the person is told how many more matching answers would settle it.
A question goes to Fine-tuning › Definitions to settle, under Floor needs a ruling, only when:
While votes are still being collected and everyone agrees so far, there is nothing there for a lead to act on. The ruling row names who voted which way, so a lead does not simply approve the bigger number. A lead presses one reading to settle it, and it can be withdrawn under Awaiting sign-off. Not sure counts towards stopping the asking, never towards settling.
Votes carry a name only when the firm signs in with Microsoft or Windows identities. On a shared access key, a vote is recorded without a name, the voter is told so, and unnamed votes can never settle a definition by themselves. On a key-only installation, floor voting still gathers opinion, but the leads settle every question.
Off by default. While off, each person sees an Ask me what words mean tick box beside the question box and a Stop asking me link under a vote, so someone who never wants to be interrupted can switch asking off for themselves. They still see which reading each answer used. Turn it on to ask everyone, with no way to opt out.
Under Answer display, off by default. It shows the Counted … as … line under qualifying answers without asking anyone to vote. It is always on while floor voting is on.
Anyone can read Business Rules. Only an administrator can save. The screen has tabs: Rules, Terms, History, Has it worked? and How this works. It opens on Rules, which has two sections, both folded until you open them, each with its count. Rules in effect by Administrator is your own policy on how to count, edited here: the list, with Always apply and Remove on each, the text editor, the save buttons, and where each rule came from. It opens by itself when you add a rule or a save needs your attention. The count is what is saved; a rule you have added but not saved is not yet in effect. Directly below it, Definitions in effect from Fine-tuning lists every definition in force, marked awaiting sign-off or confirmed: they reach the AI exactly as rules do, but they are decisions about your data settled on Fine-tuning, so they are changed only by withdrawing. A confirmed one has Withdraw it; one awaiting sign-off has Open in Fine-tuning. Save written rules and Clear written rules never touch settled definitions or terms. History holds withdrawn definitions and earlier versions of the rules; Has it worked? includes how floor voting is going.
The screen's own advice is Write definitions, not answers. A good rule defines terminology and conventions: what a term means, how to count, which records to exclude. Avoid recording conclusions. Telling the AI what it should find stops it finding anything you did not already know.
| Defines (good) | Answers (avoid) |
|---|---|
| Any status beginning 'On Hold' is held, not active. | We have 412 active files. |
| Always count distinct record identifiers after joining detail tables, never raw rows. | The Smith matter is the oldest open file. |
| Our fiscal year starts on 1 July. | Revenue this year is higher than last year. |
Limits: up to 200 rules, 400 characters each, about 30,000 characters in total. Anything too long is refused and named, not shortened for you. Edit all as text is for pasting several at once or reordering. Clear written rules removes them all; settled definitions and terms are not affected.
A rule can use wildcards: % for any run of characters, _ for exactly one. You do not have to use them; "match the name anywhere, not just the beginning" works too. Whether a match ignores capitals differs by database, so say so in the rule if capitals matter.
When Check rules against a second model is on (under AI provider), a second model plans the same question when a rule is previewed. If the two write different queries, the wording still admits more than one reading.
Open Terms, open Define a term (Administrator) (folded, and open by itself when no terms exist yet), and fill it in:
FileStatus IN ('Active', 'On Hold')) and Tables it applies to. Naming tables keeps the term out of unrelated questions; leaving it blank sends it with every question.Press Add the term, then Save terms.
History keeps the last ten versions of the rules, saved automatically before each change, with when and by whom. A save that empties the rules is never kept, so it cannot push a good version out. Restore replaces the rules on screen with that version, after you confirm. The version being replaced is kept as a backup, so a wrong restore is recoverable. Nothing in History is sent to the AI.
Go to Settings › DB Tuning › Question-SQL pairs. A pair joins a question your firm asks constantly to the query that answers it. From then on that question is answered by running your query, word for word: no AI call, no wait, and the same query every time, so the figure changes only when your data does.
In chat, an answer from a settled pair carries Verified query with the name and date of whoever settled it, and costs nothing from your AI account. Floor voting never asks on these questions. If nobody has re-checked the pair for a long time, the mark says how many days.
Settled pairs are listed under Settled questions, with the query, who settled it, and any attached wordings.
Anyone can type @ in the question box to say which tables a question is about. A short list appears as they type, and choosing one puts its name into the question. Nothing shows until the @ is typed, so it stays out of the way of people who do not want it. This is the feature most likely to be useful to you and to a DBA who knows the schema: it stops the product guessing which of three similarly named tables was meant.
Beside each table, Often needed with it offers the tables it is joined to. Those come from two places and the tooltip says which: a key your database declares, or a join that queries run here have actually used — which preparation recorded. It is never a guess about what tables are "usually" used together, because this product does not measure that.
It is a preference, not a restriction. If the question genuinely needs another table it is still used, and the answer names what it added. This is deliberate: refusing to answer teaches the user nothing, and honouring a scope that cannot answer produces a confident figure about the wrong population. There is no strict mode.
Three things worth knowing when somebody asks you about it:
Sales.invoices and Archive.invoices — typing @invoices asks which they meant rather than choosing one. Naming it in full works.On a database small enough to send whole, naming tables changes what the AI is asked to concentrate on, not what is sent — the schema travels either way. On a large one, where only the relevant part of the schema is sent with each question, it changes both.
Go to Settings › DB Tuning › Ground truth. These are figures your firm has already established as correct, from reports it already runs and trusts. The system answers each question and shows which figures it reproduced exactly. Nothing is ever edited to match an answer: if a figure and an answer disagree, that is the finding.
A figure belongs to the database it was reported from. If you switch database with unsaved drafts on screen, they are dropped and the screen says so. Saved figures can be changed in the grid and saved with Save figures.
Advanced: edit as text holds every case, including settings the grid does not show: which column carries the figure, label aliases, how close counts as equal, and whether rows your report leaves out are expected. It saves only if every case is valid, and tells you everything wrong at once.
Each question is planned, billed and recorded exactly as if someone had typed it, and counts against any daily allowance your installation has. Asking 3 times costs three times as much. The planner does not write identical SQL every time, so a single pass cannot tell a real fix from a lucky one. Use 3 or 5 after a change you want to prove. The Last check column shows each result, and a line gives how many matched exactly. Re-run after any change to your business rules.
Go to Settings › DB View Assistant. It helps you design a database view for questions that keep joining the same tables. It writes a script; it never runs it.
Shown to administrators. It lists table combinations worth a view, from recorded questions. A combination appears once at least five questions, and at least 5% of all questions, have used the same set of two or more tables; a join that had to be inferred rather than read from a declared key puts it higher on the list. Each shows the joins, whether each is declared or proved by reading rows, and where there are several child rows per parent. An empty list on a young installation is normal.
The counts come from questions asked by your Windows user on this PC, kept in %APPDATA%\EnterpriseDatabaseChat\usage-trends.json. Only the database's identifier, table names, counts and dates are kept, never question text, SQL or values; the card names the file. A colleague's questions on another PC or account are not in your counts.
Press Use these tables for a draft rather than a blank form. The recommendation knows the tables and how they can join, but not which columns, which rows or how they should be totalled, because it keeps no SQL. So the draft selects the tables, ticks the columns they join on (unless the hidden-column list could not be read, when it says so and ticks none), and offers a suggested description you adopt with Use this description or replace. It then asks two things, and Estimate & Generate View Script waits until both are answered:
When no join between the tables is known yet, there is nothing to choose from, so you write both instead: a line starting Join: in Filters and business rules, and Row grain. The button waits for those.
Your choices fill Row grain and a Join: line in Filters and business rules. Everything else is yours to add or leave. Choosing other tables drops the draft.
The Guide button offers the analyst guide for these fields.
Give the script to whoever creates objects in your database, usually the DBA. They review it and create the view with their own rights; AI CaseQuery's login is read-only and cannot. Once it exists, press I have created it — re-read the schema on the recommendation, or refresh the object list on Schema rules. The new view is then available to questions. On Schema rules, Use in the DB View Assistant takes any table or view you are inspecting into this form.
Go to Settings › Security › Import & export tuning.
The file holds your business rules and terms, settled definitions, column notes, the questions you have checked, and every previous version of your rules. It contains no provider keys, access keys, licence key or database connection, and cannot contain chat history or any key saved in a browser. It can be handed to whoever keeps your backups.
You do not have to remember to export. A copy of your rules, terms, settled definitions and column notes is saved automatically each day in the daily-exports folder beside your data. It is updated a few seconds after any change is saved, again when the application closes, and checked every ten minutes, so each day's copy is how that day ended. The last seven days are kept; older ones are removed. Each workstation, and each person on a shared terminal server, keeps its own copies in its own subfolder; the list shows, for each day, the copy taken last across all of them.
To go back to how a day ended, press Use this one beside that day. It opens the same preview as a restore from a file, and nothing changes until you press Restore these files.
Restore puts back business rules, settled definitions, column notes, the sensitive-column policy and the preparation record. Use it on a rebuilt server, or when moving from a pilot machine to the real installation.
A user clicks Not right? beside a measurement under How this was measured. In Correct this measurement they pick anything under What went wrong?, write What should it say? and Your name, and press Send for review. It is not sent to the AI and does not change the answer.
The amber Worth checking panel offers Help me sort this out…, a guided flow using Casey that ends in a definition, a correction or a handover depending on who is asking.
A copy of the whole installation folder moved to another computer keeps:
A key sealed to a workstation does not travel; move or enter the key again on that computer (checklist step 11). If a shared copy of the key has not been cleared, the copy carries it, which is one more reason to press Clear the shared copy once every workstation is done. Safeguards, the AI provider, model and thinking level start from the installation settings if they were set before the copy was made (see 3.3), otherwise at their defaults for each Windows user there; starter questions start at their defaults. To move only tuning, use an export instead (see 2.9).
The tabs across the top of Settings, in screen order. Each needs the Administrator key or role.
The ordered rollout list. Every step is described in 1.3. Return to it after any change to connections, keys or permissions and press Re-check.
For: keeping the installation running, and saying what it is licensed for.
Safe default: nothing to set. When to change: at purchase and at renewal. If you have arranged Windows groups or Entra ID already, the keys you choose here remain a fallback.
Three tabs: Who may administer, AI provider keys, Import & export tuning.
Who may administer this installation reports what is configured now. Under Change who holds Administrator and Operations Lead, choose in What to change:
| Option | Use when |
|---|---|
| A shared access key your firm sets | A small firm, or no directory. |
| A Windows group — local or Active Directory | You manage people through Windows groups, including local groups on a machine outside a domain. |
| Microsoft Entra ID — your firm's own app registration | Your IT team registers the application in your own directory. |
| Microsoft Entra ID — the app registration we publish | You use the registration we publish rather than your own. |
| Restrict use to a group, or to signed-in accounts | A different question: who may use the application at all, not who administers it. |
Fill in the fields. The screen composes the settings and shows What will be saved. Press Save to the application folder, or Copy instead if the file is already protected and someone with rights will paste it. Anything else in the file, such as a published connection, is kept. Test the new route in a second browser before you lock yourself out.
The settings a new PC or Windows user starts from until they choose their own: the AI provider, the model for each provider, the thinking level, fast mode, and nine safeguards (row limit, safe mode, question checking, explanation level, result columns, result table, the AI and query time limits, and On Floor Voting). Without them, anyone who has not chosen starts on the settings the application ships with.
They matter whenever more than one person uses the installation. On a shared network installation every PC reads them; on one PC or a terminal server every Windows user reads them; when the folder is copied to each PC, each copy keeps what it had when it was copied, so set them first. Anyone who has chosen their own setting keeps it, and other running copies pick up a change when they restart. Press Make these the installation's starting settings to make the settings in force on your PC the starting ones.
While your own settings are yours alone, because none are set or yours differ, a warning bar shows at the top of Safeguards and AI provider, with the same button and, where the installation has settings, Use the installation's settings. That resets your provider, model, thinking level, fast mode and the nine safeguards to the installation's.
Also on this tab. In Shared folder for corrections, name a folder every user can write to, typically a network folder beside the application share. Press Test & save folder; the folder is tested when you save, and the status line always shows where corrections are actually written. Leave it empty to keep corrections on each workstation. Set this before the last setup step: it is saved in the protected settings file.
Where your AI provider key is stored says which key is in force and where it comes from.
| Where | What protects it |
|---|---|
This workstation (%PROGRAMDATA%\CaseQuery\Security) | Encrypted by Windows for that computer. A copy taken elsewhere cannot be read. Anyone who may run AI CaseQuery on that computer can use it. |
An environment variable (CASEQUERY_ANTHROPIC_KEY, or CASEQUERY_OPENAI_KEY for OpenAI) | Nothing in the application folder, but the value is plain text any user of that computer can read. A compatibility option. |
deployment.json beside the application | Not used. A key found there is refused. |
Putting it in the environment instead gives the command to run in an elevated PowerShell, and notes that Group Policy or Intune can push it. Limiting what a leaked key could cost you recommends issuing the key in its own provider workspace with a monthly spend limit, so a leaked key can be revoked and could never have spent more than that cap.
Safe default: provision each workstation, then clear the shared copy. When to change: when a key is rotated, a workstation is added, or a key may have leaked.
Covered in 2.9.
For: how much data each question may return, what reaches the AI provider, how long to wait, floor voting, and what each answer shows.
| Control | Default | What it does, and when to change it |
|---|---|---|
| Safe response mode | On | Favours totals and constrains detailed listings. For databases too large to send in full, switching it off may name more unclassified columns to the AI, so classify columns in Schema rules first. |
| Confirm token estimate | Off | Requires Okay or Cancel before each AI provider call. The send button becomes Estimate & Run and a Review estimated token use dialog appears. Turn on only where every call must be approved. |
| Maximum detail rows | 1000 | Safe mode allows up to 1,000 rows. Lower it to keep listings short. |
| Control | Default | What it does |
|---|---|---|
| Check questions before sending | On | Warns the user when a question appears to contain a Social Security number, a payment card, an email address, a telephone number, or a real value from a column treated as sensitive. Nothing is sent until they decide. It reduces mistakes rather than preventing them. Leave on. |
| Also watch name columns | Off | Also catches a borrower or party name typed into a question. Surnames overlap with ordinary words, so it warns more often. Turn on if names must not travel. |
Both in seconds, from 5 to 600.
| Control | Default | What it does |
|---|---|---|
| AI provider response | 120 | Applies to the whole question. A second request after a failed query shares this budget. |
| Database query | 120 | Applies to each query. SQL Server, PostgreSQL, MySQL and Oracle cancel the statement on the server. SQLite is stopped while it reads rows, but a single long aggregate cannot be interrupted. |
Raise either only if legitimate questions time out on a slow provider or a large database.
| Control | Default | What it does |
|---|---|---|
| Ask the floor when a question is ambiguous | Off | Asks people which reading of a status word they mean. Changes how the product speaks to everybody. |
| Votes needed to settle it | 4 | Different identified people agreeing makes a provisional settled definition. |
| Stop asking after | 10 | Asked this many without agreement, it goes to the leads. |
| Everyone is asked | Off | Off lets each person switch asking off for themselves. |
How a vote becomes a settled definition folds away a short explanation. The full behaviour, including when the leads rule and why unnamed votes never settle anything, is in 2.4.
| Control | Default | What it does, and when to change it |
|---|---|---|
| Show which reading each answer used | Off | Shows a line such as Counted "open" as Open under qualifying answers, without asking anyone. Always shown while On Floor Voting is on. |
| Explanation level | Standard | Under Around each answer: Plain — answer only; Standard — adds how it was measured; Reasoning — adds the assistant's reasoning; Expert — adds the SQL it wrote. Every level shows the answer in words and any warning that an answer may be wrong, except which business rules were left out of a question: that line shows from Reasoning up. A practical ladder is Plain for executives, Standard for most staff and Reasoning for leads, who can spot a rule that should have applied and report it to an administrator, who owns the written rules. The result table follows Show result table. |
| Everyone uses this level | Off | Off lets people choose their own Explanation beside the question box, so someone checking the system can go deeper without changing what anyone else sees. On holds everyone at the level above and hides their choice. |
| Show result table | On | Displays the returned rows beneath each answer. |
| Suggest follow-up questions | Off | Offers two or three related questions under each answer, ready to click. Useful for people who do not know what to ask next. |
| Result columns | Requested columns plus supporting columns (recommended) | Right after Show result table, because it decides what that table contains. See below. |
Result columns. Supporting columns can make an answer easier to understand. Only the requested columns adds no supporting figures or joined names, and does not add sorting or rounding.
Hard protections remain active whatever is set above: writes, multiple statements, unsafe engine functions and system-catalog access are always blocked, the technical ceiling is 5,000 rows, and columns hidden in Schema rules are never sent to the AI provider.
For: the account that answers questions, the model, and how hard it thinks.
One card per provider, marked In use, Not in use or Coming later. In each:
How long the model thinks before answering. Each level shows a measured typical time per question on the connected database; your own questions will differ. Quicker is not a worse answer for most questions. The setting applies to Claude models; the screen says if the active provider ignores it.
Fast mode, off by default: the same model generating faster, billed at roughly double. Your provider organisation must be granted it; if not, questions run at normal speed and the screen says so.
Check rules against a second model: before a rule is recorded, a second model plans the same question. If both write the same query, the rule says one thing. If they differ, the wording still admits more than one reading. It runs only when a rule is previewed, never on an ordinary question, at one extra call each time. The second model is chosen for you from the answering model; where none is paired, the switch says so.
Safe default: the shipped model, default speed, fast mode off. When to change: turn on the second-model check while rules are being written in earnest.
The example questions on the chat screen before anyone has asked anything. They are presentation only and never reach the AI, so word them in your firm's own vocabulary. Type One question per line, then Save questions. Restore the supplied set puts the originals back. Starter questions are saved for the Windows account that saves them, on that computer.
Casey is the guide to this product: it explains and drafts, changes nothing and never runs a query. It uses your AI provider account. Full detail is in Working with Casey.
| Control | Default | What it does |
|---|---|---|
| Offer Casey on the chat screen | On | Everyone gets the Ask Casey tick box beside the question box and the Ask Casey launcher (Ctrl+K). |
| Let Casey read what your firm has written down | On | Lets Casey read your terms, business rules, settled definitions and column notes, so it can answer "why is my rule not firing" rather than only "how does the Rules screen work". Still no rows from your database, and never notes on hidden columns, though a column note can name the values a column holds. |
| Starter questions for the assistant | Supplied set | Casey's own suggested questions, one per line. |
Press Save. Casey accepts questions up to 2,000 characters, counts towards any daily limit, and refuses if its product guide does not match the running version after an update.
Every supported database is opened read-only, and the driver for each type ships with the application.
A QuickBooks company file has no SQL behind it, so AI CaseQuery imports a snapshot it can query.
Whatever is connected now is available only on this machine until it is published. Publish this connection hands it to every workstation, so nobody else configures anything. Stop sharing removes it. Do this before the last setup step: once the settings file is locked down, only an account allowed to change it can publish.
Connections saved on this computer, to reopen with a click.
Combines answers across several databases. It needs a licence that includes it; otherwise the card says the installation is licensed for one database.
If the list is empty although you saved a second connection, you are probably connected to that second database. Connect to your main database and it appears.
Appears when combining databases. For each database, including the connected one, say which column holds the firm's case number, the status and the rest. Nothing is assumed: a field bound to the wrong column gives answers that look ordinary and are about something else. A database not yet described cannot be asked across. If a business rule changes after a value was mapped, the card warns, and questions across databases are not totalled until you confirm the mapping again.
A database is described once for each kind of record it holds, chosen with the buttons at the top of the card: Matter, Invoice, Bill, Time entry. The firm's accounting system is usually described for its invoices; a case system that raises its own invoices can be described for both. On anything but a matter the first thing asked for is the Match field — the column that identifies the case that record belongs to, which in QuickBooks is usually inside the Customer:Job name. Without it that database's records can still be totalled on their own, but nothing can be answered per case.
Describing a kind of record is not the same as being able to ask about it. Connect counts and lists records; it does not yet add up the money on them, so invoices, bills and time entries are described for the sake of linking a case to its money, and questions about them are refused until that arrives.
Two systems can write the same case number differently — FC-2025-00123 in one and 25-123 in the other — and matched exactly they are different cases. This panel, inside each described database, is where you say how that database writes it: the layout, whether the value is the whole column or the part after the last colon (which is how QuickBooks writes Customer:Job), whether to ignore leading zeros, and how to read two-digit years. Check it reads every value on both sides and reports what would match, what could not be read, near misses and anything that would turn two cases into one. Nothing is saved until that check comes back clean, and it is run again on saving rather than trusted from a moment ago.
Safe default: one read-only connection, published. When to change: a new database, a password change, or a licensed second database.
Three tabs, all described in Part 2:
For: browsing the schema, hiding columns, writing column notes, and choosing what opens when someone clicks a case number.
Search tables, views or columns, filter to Views or Tables, and refresh with the ↻ button after the database changes. Select an object to see Columns, Sample data, Definition and Relationships.
Tick a column only if nobody should see its values: a Social Security number, a bank account or card number, a password. A ticked column is never sent to the AI, never displayed to anyone, and any query naming it is refused.
Do not tick a column just because it looks personal. Party names, addresses and reference numbers are how staff and the AI identify records; hiding them stops ordinary questions working.
Safe default: hide nothing unless it is truly never needed, but save the policy either way. When to change: before preparing, and whenever a sensitive column is added. After lock-down, change it as an account allowed to write the policy file.
No database rows or query results are sent to the AI whatever you tick. A value written into a rule or column note, including code lists written by preparation, travels as part of that instruction.
Explain abbreviated or coded columns so the AI reads them correctly. Each visible column has a note box; hidden columns do not, because a note on them could never be read. The values a column holds are shown with counts. A full list short enough to fit can go into the note; when there are too many values, only the most common are shown and not offered, because a partial list would tell the AI the rest do not exist.
Show only columns needing a note narrows the list. Press Save notes.
What opens when someone clicks a case number in a result. Tick fields in the order you want them read; ticking adds a field to the end, so re-ticking moves it. Use what the primer suggests fills it from preparation. Press Save. Hidden columns never appear, whatever is ticked.
Designing a reviewable CREATE VIEW script from recorded questions or from sources you choose. Described in 2.8.
Records what each question sent to the AI provider, so you can show an auditor what left the machine. Query results are never sent, so the log never holds them. It holds your schema, column notes and questions, and a database's own error text, which can quote a value, so handle it as confidential.
Gathers what support needs into one file you can read before sending. Nothing is transmitted from here. It contains the questions people asked, the shape of your schema, an extract of the activity record with names replaced by user-1, user-2, and the application's own log. It never contains an access key, a provider key, a database password, a licence key or a row of your data.
Safe default: both off. When to change: an audit request, or a fault to report.
| Screen | Anyone (no key) | Lead | Administrator |
|---|---|---|---|
| Chat | Ask questions, vote on the floor, send corrections, use Casey, personal choices | The same | The same |
| Fine-tuning › Definitions to settle | Read | Settle or set aside definitions; rule on Floor needs a ruling; Send to the DBA | Everything a lead can do |
| Fine-tuning › Baseline questions | Read | Run sections, mark answers, add questions | Everything a lead can do |
| Fine-tuning › Awaiting sign-off | Read | Read | Confirm this / Withdraw it |
| Writing a definition | — | Check the wording, See what this changes | The same |
| Business Rules | Read | Read | Save rules and terms, Restore from History, accept a drafted rule |
| Settings (every tab) | — | — | Everything, including preparing the database, Question-SQL pairs, ground truth, the DB View Assistant, and export and restore |
An Administrator reaches every screen. A person's name is asked before any recorded decision. An unlock lasts until the browser tab closes.
Every call below is billed to your firm's own provider account and may count towards a daily question limit if your installation has one.
| Activity | Cost |
|---|---|
| An ordinary question in chat | Provider calls per question. A settled Verified query costs nothing. |
| Preparing the database, including reading system documentation | Uses the provider. Table data is not sent. |
| A baseline section run (Answer this section) | Each question is asked twice: up to 2 calls per question, at most 3 if a repair is needed. Questions already verified cost nothing. |
| Ground truth Check the answers | Each question is asked once, 3 or 5 times: the cost multiplies accordingly. The screen states the number before you confirm. |
| Casey, including Help me sort this out… | Uses the provider for each question. |
| Rule and definition checking: Check the wording, See what this changes | Uses the provider. Check rules against a second model adds one extra call each time. |
| Request SQL for this question on Question-SQL pairs | Uses the provider. Run it and show me runs your query against the database only. |
| View generation (Estimate & Generate View Script) | Uses the provider. Recommendations and the read-only preview do not. |
| Fast mode | Roughly double the price per call, where granted. |
| Floor voting, starter questions, Schema rules, exports | No provider calls. |
For installation problems (Node.js, extraction, ports, updates, provider refusals, database connection errors), see the Troubleshooting section of the Installation Guide (INSTALL.md).
| What you see | Likely cause, and what to do |
|---|---|
| Settings asks for the key again | The unlock ended: the browser tab closed or the application restarted. Unlock again. |
| The evaluation keys no longer work | A licence was applied. Use the keys chosen then, or your group or Entra role. If nothing was arranged, the application writes a settings file named deployment.json with placeholder values: replace them with keys of your own, save and restart. |
| Signed in with a work account but nothing can be configured | The account holds no role. Assign it in the Microsoft admin centre (checklist step 3). |
| Added someone to a Windows group but they have no access | They must sign out of Windows and back in. |
| A checklist step will not tick | Steps marked Checked cannot be ticked by hand. Do the work, then press Re-check. Step 16 names the exact group and file still at fault. |
| A workstation is refused an AI key | The key was not moved onto that computer, or the shared copy was cleared first. Run Move the key onto this computer there, or provision it as the Installation Guide describes. |
| One person's answers use a different key or model | A personal key under Just for you, on this PC outranks the firm's. Check Which key gets used on their computer. |
| The database looks smaller than it is | The login lacks rights. See What rights does the login need? |
| The preparation badge reads Schema changed — run again | Tables or columns changed. Prepare again. |
| A settled question is no longer answered as a Verified query | The database changed since it was written, so it is not used. Check the query and settle it again. |
| Floor voting never asks anyone | It was switched on for a different Windows account or computer, the question did not meet both conditions in 2.4, or the person switched asking off. |
| Nothing is waiting under Floor needs a ruling | Normal while votes are still arriving and everyone agrees so far. |
| Floor votes never settle a definition | The installation signs in with a shared key, so votes carry no name. The leads settle instead. |
| A rule was refused on save | Too long (over 400 characters) or over the total. The refused text is kept on screen; shorten it or split it. |
| Rules disappeared | Someone saved an empty or different set, or restored an export. Use History › Restore, or restore your last export. |
| Corrections are not reaching the shared folder | Check the status line under Shared folder for corrections. A laptop off the network writes locally and the entry says so. |
| A new view is not available to questions | The schema has not been re-read. Press I have created it — re-read the schema or refresh on Schema rules. |
| Answers to the same question still vary | Look at Has it worked?, settle the definition behind it, or check with ground truth asked 3 times. Settle it as a Question-SQL pair if it must never vary. |
| Casey refuses after an update | Its product guide does not match the running version. Confirm the update completed on this installation. |
| Something else | Switch on Debug mode, reproduce it, and create a support file (see 3.11). |