Guides

Setting Up the Compendium

Start your journey on the stage

Setting Up the Compendium

The Compendium is RoleCall's AI-managed knowledge layer. Instead of relying on keyword triggers to fire when the right word happens to appear, a dedicated AI sidecar reads each scene, pulls the entries that matter into the narrator's context, writes new entries when something important happens, and keeps the whole thing organized as the story grows.

This guide walks you from a fresh chat to a working Compendium-enabled scene. By the end you'll have it turned on, a lorebook attached, a tree built, and you'll know what to watch for as the AI starts filling in pages on its own.

A note on naming. You'll sometimes see this called TunnelVision or TV in older menus and slash commands — it's the same system, renamed for clarity. The slash command prefix is still /tv. The reference doc is Compendium.

1. Where the Compendium Lives

Open any scene and find the Compendium entry on the chat rail — same surface that holds your other wings (Cast, World, Prompting, and so on). Tap or click to open it.

The wing has five sub-tabs running across the top: Dashboard, Browse, Storyboards, Activity, and Settings. You'll spend most of your time on Dashboard (status at a glance) and Browse (read and edit the pages the AI has written). Activity is where you watch the AI work in real time. Settings is where you tune behavior.

If you've attached more than one Compendium-managed lorebook to the chat, a dropdown in the header lets you switch between them. Every tab below re-mounts pointed at the selected book.

2. Turn It On for a Scene

The Compendium is per-chat. You decide whether it runs in this scene, not globally — so you can experiment freely without committing your whole library to it.

Two switches matter:

The master toggle. Lives in your scene's settings (Quick Play and Stage). When this is off, no sidecar passes run at all.
TV Managed on a lorebook. A book has to be flagged "TV Managed" before the sidecar will read from or write to it. Books without this flag are invisible to the Compendium.

You can turn on the master toggle and have zero managed books — the sidecar will still run pre-gen retrieval (and find nothing), and post-gen will refuse to write because it has nowhere to write to. So the next step is attaching a book.

3. Attach a Lorebook (or Start Fresh)

The Compendium needs something to manage. Two paths.

Start Fresh

Create an empty lorebook in your library, attach it to the chat, and toggle TV Managed on. The Compendium will build it from nothing as you play — every named character, every visited place, every faction the narrator mentions becomes a typed entry over time.

This is the lighter setup. It works well when you're starting a new story and don't want to seed any prior assumptions.

Attach an Existing Lorebook

If you already have a lorebook for this world — yours or one you've forked — attach it to the chat the way you normally would. Then toggle TV Managed on for that book.

You see this	What it means
Copy & Enable	Forks the lorebook into your library, marks the fork as TV-managed, and leaves the original untouched. Safe default for community books.
Use Original	Enables the Compendium directly on your own lorebook. No safety copy. Use this when you own the book and want edits to stick to it.
Fork & Enable	Shown when the source book isn't yours. Forks and attaches in one step.

The Compendium needs write access to attach. If you're working with someone else's book, the fork pattern keeps the original creator's version pristine while you build out your own copy.

4. Build the Tree

The Compendium organizes entries into a navigable tree — Characters here, Locations there, Organizations and Items in their own buckets. Building the tree once gives the sidecar a map to walk when it picks context.

Open the Compendium wing's Dashboard and click Build Tree. You have two methods:

Method	What it does	When to use it
From Metadata	Auto-clusters entries into semantic groups using keyword analysis. Fast. No AI cost.	Most lorebooks. The default for a reason.
With LLM	Sends entry titles and keywords to your AI for intelligent categorization. Slower. Costs tokens. Produces better groupings on messy unstructured books.	Large community books with hundreds of unstructured entries.

A progress drawer streams the build as it runs. You can dismiss it; the build continues in the background. When it finishes, the TV Tree sub-tab under Browse fills in with the new layout.

You can rebuild any time. The slash command /tv rebuild triggers an LLM rebuild from chat. Any locks or "forbidden" flags you've set on branches survive rebuilds — your hand-curated structure won't get wiped.

5. Your First Run

That's the setup. The Compendium runs automatically each turn from here.

Send a message. Watch the Activity tab. You'll see a few things happen in sequence:

Pre-Gen pass — the sidecar reads recent messages, walks the tree, picks which entries belong in the narrator's context. You'll see rows tagged PRE with actions like "Searched Locations (3 entries)".
The narrator writes the actual response. This is the main message you read in chat.
Post-Gen pass — the sidecar reviews what just happened and updates the Compendium. New characters get typed entries. Existing entries get filled in. Soft-deletes for things that died or got destroyed. You'll see rows tagged POST.
Historian — periodically, the historian writes a structured summary of the recent window. You'll see rows tagged HIST.

The pages don't always show up after a single message. Real story facts emerge over a few turns. Give it five to ten exchanges before judging whether the Compendium is doing useful work.

6. Reading and Editing Pages

Switch to the Browse sub-tab. On the left, you'll see entries grouped one of three ways:

By Type — typed buckets: Characters / Locations / Items / Organizations / Events / Powers / Threats / Worldbuilding / Creatures / Plots.
TV Tree — the AI-built hierarchical tree, drag-and-droppable.
Summaries — just the historian's summary entries, in narrative order.

Click any entry to open the detail pane. You'll see structured typed fields — a Character page has tagline, affiliations, arc notes, relationships, hidden knowledge; a Location page has region, danger level, connected places, notable inhabitants; and so on.

Every field is editable inline. Click, type, save. Three controls per entry worth knowing:

Control	What it does
Lock	The AI is no longer allowed to overwrite this entry's content on future post-gen passes. Use when you've hand-edited something the AI keeps trying to "correct."
Forbidden	This entry must not exist. On the next reorganize, the AI redistributes its content elsewhere and removes the entry. Useful for clearing out hallucinations.
Soft delete	The entry is hidden but recoverable. The AI is told it no longer exists. Use this when a character dies or a place is destroyed — the story acknowledges it without losing the history.

The relationship graph in the detail pane is clickable — jump from one character to a connected one without going back to the list.

See the full field reference at Compendium.

7. Standardization

If you attached an existing lorebook with entries written before the Compendium era, those entries are classic — plain text content with optional keywords. They work, but the AI can't read their structured fields because they don't have any.

Standardize is a one-shot pass that converts classic entries into typed entries. Open the Dashboard and click Start Standardization. The AI walks every entry, classifies it into the right type, fills in the structured fields for that type, and stores the original content so the whole run is fully reversible.

You'll watch a live progress drawer. Batch size, model, and whether to auto-rebuild the tree after are configurable in Settings → Standardization Defaults.

If something goes wrong, Revert Last Standardization (also in Settings) restores every entry to its original content and removes anything the run split off. Confirms before firing.

When to standardize:

After importing a lorebook from another platform.
After attaching a community book and you want the dossiers to recognize what's in it.
Once — at the start — for any book that pre-dates typed entries.

When not to standardize:

The book is already typed. The standardizer skips already-typed entries, so you're spending tokens for nothing.
You're mid-scene and the run might rewrite an entry the narrator is about to read. Wait for a natural break.

8. The Activity Log

Activity is your window into what the AI is actually doing. Every sidecar action gets a row.

The floating popup contains the complete latest generation. Use Open full Activity to open the Compendium wing directly on the same feed. Trigger rows name the exact entry and lorebook, show whether it was constant (Always on), keyword-matched, or retrieved by the sidecar, and keep a separate Why this entry explanation for each result.

With Prompt Inspector enabled, a normal turn can contain both Compendium Retrieval before narration and Compendium Capture after narration. They are separate calls with separate jobs. Retrieval can also retry when an answer is unusable or below the configured minimum. The inspector labels the phase, purpose, and retry round so you can tell exactly which call you are reviewing.

Each row shows:

A phase badge — PRE, LIVE, POST, or HIST — so you can tell which pass produced it.
A model tag — which model handled the call.
An expandable Why? dropdown with the AI's reasoning.
An expandable Details dropdown with the full structured payload.

Filter the feed by All, Entries (only writes), or Tools (only reads). Pop the feed out into a floating window if you want to watch the AI work while you keep chatting.

This is where you catch hallucinated facts before they cement themselves into your world. If you see a Remembered row for a character that doesn't exist, expand Why?, then go to Browse and either edit or forbid the entry.

9. Settings Worth Tuning

The defaults work for most stories. Once you've run a few scenes, these are the knobs most worth touching.

Pre-Generation

Setting	Default	When to change
Max entries to retrieve	10	Bump up if your scenes are losing context. Bump down if pre-gen feels heavy.
Minimum entries to retrieve	0	Set to 3 – 5 if the sidecar keeps under-retrieving in long-running stories.
Entry detail level	Names + Taglines	Switch to Names + Taglines + Preview if the sidecar is picking the wrong entries.
Context messages	20	Bump up if recent events keep getting forgotten in retrieval.

Post-Generation

Setting	Default	When to change
Write frequency	1	Set to 2 or 3 on long sessions where post-gen feels expensive — runs every Nth turn.
Mandatory tool calls	On	Turn off if the model keeps inventing busywork on quiet turns.

Historian

Setting	Default	When to change
Interval (messages)	0 (off)	Turn on and set to 15 – 20 for tightly paced stories. Push to 30 – 40 if summaries feel redundant. 0 disables the historian entirely.
Summary-aware context truncation	Off	Turn on for very long sessions — once a window is summarized, the raw messages drop out of context, freeing tokens. The summary stays.

Advanced — Classifier Model

Tucked at the bottom of Settings. Pre-gen retrieval, tree-building, and standardization are all bucket-picking tasks — they don't write prose. Pointing them at a cheaper classifier model (Haiku, GPT-4o-mini, Gemini Flash) cuts those passes' cost dramatically without affecting narration quality. Highly recommended once you've seen a few token bills.

10. Troubleshooting

"The AI keeps inventing new pages for the same character." Standardization is misbehaving, or the sidecar isn't seeing existing pages in pre-gen. Try: bump Entry detail level to "Names + Taglines + Preview" so the sidecar can tell the duplicates apart. Then run Reorganize from the dashboard to consolidate.

"Pages aren't appearing at all." Three things to check, in order:

Is the master Compendium toggle on for this chat?
Is at least one attached lorebook marked TV Managed?
Does your model support tool calls? Some smaller open-source models can't reliably call structured tools, which the Compendium relies on.

"A page got overwritten with garbage." Edit it back to what you want, then Lock the page. Locked entries survive future post-gen passes untouched.

"The activity feed shows errors." Click the error chip on the row for the full message. Most errors fall into one of three buckets: the model failed mid-tool-call (retry the turn), the model exceeded its context window (lower Context messages or pick a model with a bigger window), or the entry hit a hard schema constraint (rare, surfaces with a clear message).

"My lorebook is read-only." That means the book's AI Permission is set to Read Only. The sidecar will retrieve from it but never write. Change it in Settings → Managed Books if you want writes.

What to Read Next

Compendium reference — the full field-by-field rundown, including every slash command and storyboard dossier.
Lorebooks — the upstream source. How keyword entries work, how to write good ones, and how they interact with the Compendium.
Scenes & Chat — where the master Compendium toggle lives, and how settings cascade from your chat into every pass.
Story Director — the dossier system that consumes Compendium data and surfaces it as Quest Board, Cast, Calendar, Map, and Renown views.

The Compendium gets dramatically better with use. Run a few scenes, watch the Activity feed, lock the pages you care about, forbid the ones you don't. By scene five the AI will know your world about as well as you do.