Injection and Context Control

Injection and context control

How and where entries land in the prompt, and the per-chat overrides that let you steer what gets injected on this chat without touching frontmatter.

For the upstream stages that decide which entries reach injection, see Pipeline and Entry Matching and Behavior.

---

Per-entry injection position

Entries can override the global injection position via frontmatter:

Field	Values	Description
position	before, after, in_chat	Where to inject
depth	number	Chat depth (for in_chat)
role	system, user, assistant	Message role (for in_chat)

Entries are grouped by their effective position (global default or override). Each group is injected separately via setExtensionPrompt.

Example: most lore at depth 4 as system messages, but a character's dialogue hints at depth 1 as user messages.

See Writing Vault Entries for templates.

---

Prompt Manager integration

Set Injection Mode to Prompt List to register DLE's injections as named entries in SillyTavern's Prompt Manager. This lets you drag them to any position in the prompt order: before character definition, after Author's Note, between example messages, wherever.

How it works:

1. Switch injection mode to "Prompt List" in DLE settings.
2. Generate at least once so the entries appear.
3. Open the Prompt Manager and find deeplore_constants and deeplore_lore.
4. Drag them to your desired position, or switch to Absolute mode with a custom depth.

Notes:

• Requires a Chat Completion API (OpenAI-compatible).
• Per-entry frontmatter overrides with custom position/depth still create separate injection groups.
• The deeplore_notebook (Author's Notebook) and deeplore_ai_notepad (AI Notepad) entries also appear in the Prompt Manager when those features are enabled, so you can position them independently.

---

Author's Notebook

A persistent per-chat scratchpad injected into every generation. Use it for author notes, scene direction, tone guidance, or anything you want the writing AI to always see for this specific chat.

How it works:

1. Open the notebook via /dle-notebook or the "Open Notebook" button in settings.
2. Write any text. It's saved per-chat in chat_metadata.
3. The notebook content is injected into every generation as a separate prompt, independent of the entry pipeline.

Setup:

1. Enable "Enable Author's Notebook" in Settings Reference.
2. Choose injection position (Before Main Prompt, After Main Prompt, or In-chat at depth).
3. Open the editor and start writing.

Notes:

• Notebook content is stored in chat_metadata.deeplore_notebook and persists across sessions for that chat.
• Injection is independent of the lorebook pipeline. The notebook always injects when enabled, regardless of entry matching.
• Has its own position, depth, and role settings separate from the main lorebook injection.

---

Per-chat pin/block

Pin entries to always inject, or block entries from injecting, on a per-chat basis. Pins and blocks are stored in chat_metadata and survive page reloads.

Commands:

• /dle-pin <entry name>: pin an entry (always inject in this chat).
• /dle-unpin <entry name>: remove a pin.
• /dle-block <entry name>: block an entry (never inject in this chat).
• /dle-unblock <entry name>: remove a block.
• /dle-pins: show all pins and blocks for the current chat.

How it works:

• Pinned entries are force-injected like constants, regardless of keywords or AI selection. They get priority=10 to give them the best shot at surviving budget truncation.
• Blocked entries are removed from the pipeline before injection, regardless of matches. Blocks override constants: a blocked constant is removed.
• Pins and blocks apply after the main pipeline runs but before formatting.

UI access: the Drawer Browse tab has inline pin/block buttons on each entry. No slash commands needed. Per-chat injection counts are displayed as badges on each entry.

---

Contextual gating

Filter entries based on the current story context using gating fields in frontmatter. DLE ships with four default fields (era, location, scene_type, character_present). You can define additional custom fields via the rule builder, accessible from the Drawer Filters tab toolbar (gear icon) or the "Edit Fields" button in the Settings popup.

Default frontmatter fields:

Field	Type	Description
era	string | string[]	Entry only injects when the active era matches any value
location	string | string[]	Entry only injects when the active location matches any value
scene_type	string | string[]	Entry only injects when the active scene type matches any value
character_present	string[]	Entry only injects when any listed character is present

Custom fields work the same way: add the field name to your entry's frontmatter with one or more values, then set the active value via /dle-set-field. Field definitions are stored in DeepLore/field-definitions.yaml in your vault. Each field definition specifies a type (string, number, boolean), a multi-value flag, a gating operator (match_any, match_all, not_any, exists, not_exists, eq, gt, lt), and a tolerance level (strict, moderate, lenient).

Tolerance controls what happens when an entry has a value for a field but the active context doesn't:

• Strict: entry is blocked (out of context).
• Moderate: entry passes through.
• Lenient: like moderate, plus match_any and match_all non-matches also pass. Precision operators (eq, gt, lt, not_any) always filter.

Commands:

• /dle-set-field <name> [value]: set any gating field (built-in or custom). With no value, opens a browse-and-select popup.
• /dle-clear-field <name>: clear a gating field from the active context.
• /dle-set-era [era]: alias for /dle-set-field era. With no argument, opens a browse-and-select popup.
• /dle-set-location [location]: alias for /dle-set-field location. With no argument, shows a browse-and-select popup.
• /dle-set-scene [type]: alias for /dle-set-field scene_type. With no argument, shows a browse-and-select popup.
• /dle-set-characters <names>: alias for /dle-set-field character_present. Comma-separated list.
• /dle-context-state: show all active gating fields (built-in and custom).

Notes:

• Context state is stored per-chat in chat_metadata.deeplore_context.
• Entries without contextual fields are unaffected (always pass through).
• Partial matches work: an entry with only era set is filtered only on era, regardless of other fields.
• Custom fields appear automatically in the Drawer Browse tab as filter dropdowns and in the Filters tab with status dots and impact counts.
• The relationship graph can color nodes by any custom gating field.
• Force-injected entries (constants, seeds, bootstraps, pins) are exempt from contextual gating.

---

Per-chat folder filter

Restrict injection to entries from specific vault folders. When a folder filter is active, only entries whose vault folder path matches the filter are eligible for injection. Useful for vaults organized by storyline, setting, or campaign.

Commands:

• /dle-set-folder [path]: set the folder filter. With no argument, opens a browse-and-select popup showing all folders in the vault.
• /dle-clear-folder: clear the folder filter so entries from all folders pass.

How it works:

• Folder filtering runs as both a pre-filter (before AI search, so the AI doesn't waste selections on filtered entries) and a post-filter (after selection, as the authoritative gate).
• Constants, seeds, bootstrap entries, and pinned entries are exempt.
• The filter is stored per-chat in chat_metadata.deeplore_folder_filter.
• Multiple folders can be selected. The filter uses OR logic: entries from any selected folder pass.
• Subfolder matching: an entry under Lore/Magic/Schools matches a filter for Lore/Magic.

---

Confidence-gated budget

AI search over-requests entries (2x your configured max), then sorts results by confidence tier (high, medium, low) before applying the budget cap. High-confidence picks are prioritized. When the budget is tight, the most relevant entries make the cut.

How it works:

1. The AI is asked to select up to 2x your maxEntries setting.
2. Results are sorted: high confidence first, then medium, then low.
3. The budget and max entries caps are applied to this sorted list.
4. Low-confidence picks are the first to be cut when the budget is tight.

The AI gets room to suggest marginal entries without displacing clearly relevant ones.