Settings¶
Open settings from the gear icon in the top bar. The menu shows your display name and organization, then these items:
- Settings — the main Settings screen (connection, models, display)
- Organization Settings — rename your organization (server mode only, owner only)
- Setup Wizard — re-run the first-run setup flow at any time
- Logout — sign out of your account (server mode only)
The Settings screen is one scrollable page organized into sections. This page documents each section in the same order you see on screen.
Connection¶
Connection settings control whether Recall runs entirely on this computer or talks to a shared server.
| Mode | What it means |
|---|---|
| Entirely Local | The Recall Core service runs on your computer. All documents, matters, and AI models live on this machine. No internet required after setup. |
| Backend from Server | The Flutter app connects to a shared Recall server over HTTPS. Documents and matters live on the server. You need an account and an invite to register. |
Server URL (Backend from Server mode only). Must be HTTPS. Recall rejects non-HTTPS URLs; this is not configurable.
Test Connection. Sends a small health check with a 5-second timeout. Shows success, a timeout message, or a targeted error for DNS, TLS, or URL format problems.
Switching between modes takes effect immediately — no restart required. If you save a server URL, it is stored in connection_settings.json inside your app support directory.
Accounts and Organizations¶
In Entirely Local mode there is no account UI; everything is owned by the built-in local-user and a single local-org. In Backend from Server mode Recall requires a login.
Register¶
On the login screen, tap Register:
- Enter your Display Name, Email, and Password.
- Enter the Invite PIN your administrator sent you. Registration is invite-gated on every server; without a valid, unexpired PIN you cannot create an account.
- Optionally enter an Organization Name. If the invite is tied to an existing organization, your account joins that organization and the name field is ignored. Otherwise a new organization is created with you as owner.
Log in¶
Enter your email and password on the login screen. Tokens last 24 hours; when a token expires you are prompted to log in again. Your last-used email is remembered locally in connection_settings.json.
Organization Settings¶
Open the gear menu → Organization Settings to rename your organization. Only the organization owner can do this. The rename changes the display name and slug; the organization ID is stable, so existing sessions stay valid.
Organizations are the unit of data isolation. Each organization has its own fully isolated storage on the server. Data from one organization is never visible to another.
Logout¶
Gear menu → Logout. Clears the token from local storage and returns you to the login screen.
Word Integration¶
On macOS, Word integration has two modes:
- Recommended — Apple Events. Recall talks to Word through the macOS Apple Events bridge. Works with any Word installation. The first time you use it, macOS shows an Automation permission prompt; you must allow Recall to control Microsoft Word.
- Advanced — Word Add-in. Recall runs a local HTTPS server and Word connects to it through a manifest-based task pane add-in. Useful if Apple Events is blocked or doesn't work on your Mac.
macOS-only controls in this section:
- Request Word Permission triggers the Automation prompt explicitly. Use this if the prompt never appeared or you accidentally denied it.
- Test Connection runs a minimal health check against Word and shows the result (pass/fail plus a short log line).
On Windows, Word integration uses COM automation through the Recall Word Bridge. It starts automatically when the app launches; there are no mode or permission controls. If Word isn't open, the status indicator shows disconnected — just open Word and it reconnects within a few seconds.
See Word Integration for the full workflow.
Semantic Search¶
Choose the embedding model Recall uses for semantic search. Changing the embedder invalidates the current index, so you'll be prompted to rebuild — this re-processes every document in every matter and can take minutes.
| Embedder | Size | Notes |
|---|---|---|
| ModernBERT | ~570 MB | Bundled with the full install. Legal-domain fine-tuned. |
| E5-Large-Instruct | ~1.4 GB | Optional download. Multilingual, general-purpose. |
| all-MiniLM-L6-v2 | ~90 MB | Optional download. Smallest/fastest; lower quality. |
| Custom | varies | Drop a .gguf file into the custom models folder (see below) to make it selectable. |
ModernBERT and E5 produce higher-dimensional vectors than all-MiniLM, which contributes to better retrieval quality at the cost of more disk space and slightly slower indexing.
Generation¶
The generator model is used by Clerk (Edit, Assist, parenthetical generation) and by any feature that needs text output from an LLM.
| Generator | Size | Notes |
|---|---|---|
| Llama 3.1 8B Instruct | ~4.7 GB | Bundled with the full install. Good general-purpose quality. |
| Phi-3.5 Mini Instruct | ~3.8 GB | Optional download. Smaller, faster, sometimes prefers tighter answers. |
| Disabled | — | Turns generation off entirely. Clerk features show a disabled message. |
| Custom | varies | Drop a .gguf file into the custom models folder to make it selectable. |
In server deployments an administrator can configure Recall Core to use a remote OpenAI-compatible endpoint (vLLM etc.) instead of a local generator. In that mode this setting is read-only.
Enhancements¶
Advanced retrieval features. Most are off by default; flip them on if you want to experiment.
- Query Expansion. Off by default. Uses the generator to rewrite your query into alternatives before searching; can improve recall at the cost of latency. Requires a generator model.
- Reranking. Off by default. Re-orders search results using the Core ML GTE Reranker on macOS. Not available on Windows (Core ML is required).
- Retrieval LLM (ChatQA). Off by default. Optional retrieval planner that runs ChatQA-2 locally for query planning. Two quantizations are downloadable (Static Q4_K_M and iMatrix Q4_K_M). The toggle is locked off until at least one variant is downloaded.
- Corpus Search. On by default. Enables a backend knowledge corpus that supports query expansion, editing, and drafting. Uses a fixed ModernBERT embedder against
CorpusVec.sqlite. Requires Core ML, so it's effectively macOS-only — and is not available in server mode today.
Display¶
- Content Font Size. Font size for pin excerpts, Clerk output, and other body text. Adjustable 10–24 pt.
- Theme. Dark or light. Preference is saved to
display_settings.json.
Custom Models¶
You can add your own GGUF files. Recall scans the custom models folder on startup and exposes anything it finds as a selectable embedder or generator.
Drop .gguf files into:
Then click Refresh Custom Models to re-scan. For data-root paths, see Data locations below.
Matter-specific settings¶
Matter settings live on the matter itself, not on the Settings screen. Open the matter picker and use the edit action.
The matter edit dialog exposes:
| Field | Values |
|---|---|
| Matter name | free text |
| Citation Style | federal or california |
That's it — there is no jurisdiction, practice area, or court selector. Citation style determines how generated citations are formatted (see Citations).
Each matter also stores a notes markdown field — but that's the content of the Write-mode Notes editor, not something you set from a matter dialog.
Status indicators¶
The status pill next to the matter picker in the top bar polls Recall Core every 3 seconds. Click it to open the popover with details:
| Indicator | What it checks |
|---|---|
| Server | /health responds successfully |
| Word | Word is open and the Word Bridge has a live connection |
| Embedder | The selected embedding model is loaded and ready |
| Generator | The selected generator is loaded (or marked Disabled) |
If any indicator is red, the popover explains why. The Troubleshooting page covers the common fixes.
Auto-Updates¶
Recall checks for updates in the background, 30 seconds after launch and then every 4 hours. Updates are never downloaded or applied automatically — when one is available, a banner appears and you decide when to proceed.
See Auto-Updates for the full behavior.
Data locations¶
Recall stores data under a single root directory per installation.
| Platform | Data root |
|---|---|
| Windows | %LocalAppData%\Recall\ (i.e. C:\Users\<user>\AppData\Local\Recall\) |
| macOS | ~/Library/Application Support/Recall/ |
| Server | server-configured (set by your administrator) |
Inside the data root you'll find subdirectories for organization data (one per organization), downloaded models, custom models you've added, and the local-mode HTTPS certificate. In local mode there is a single built-in organization; in server mode the server creates one per registered organization.
Connection and display preferences live in the app support directory alongside the data root — back those up if you want to preserve your server URL and theme choices.
Re-running setup¶
Gear menu → Setup Wizard reopens the setup flow from the first step. Nothing you've done is wiped — it's just a convenient way to revisit model downloads, Word integration, or the create-first-matter step.