02.10: Sample Dialogues

Few-shot example dialogues to anchor the LLM’s voice.

File: src/utils/text/context/templates.ts:199-278

Mission

Emit paired user/model context items from the persona’s sample_dialogues_in and sample_dialogues_out arrays. These act as few-shot examples — the LLM sees ${user}: question followed by ${botName}: answer, primes its persona voice. Skipped on impersonation (samples are persona-flavored output examples) and when arrays are empty or mismatched in length.

Input

• client, guildId
• triggererName, botName
• tomoriState (provides sample_dialogues_in, sample_dialogues_out)
• tomoriConfig.humanizer_degree, tomoriConfig.personal_memories_enabled
• isUserImpersonation
• uncensorInputOptions: { unicodeSpacesEnabled, sanitizeEnabled }
• convertMentions

Output

Promise<StructuredContextItem[]> — up to 2 × sample_dialogues_in.length items (one user + one model per pair, or just model if the user side is UNPAIRED_SAMPLE_DIALOGUE_SENTINEL). Each tagged DIALOGUE_SAMPLE.

Side effects

• Humanizer transform — when humanizer_degree >= HumanizerDegree.HEAVY, both user and model sample text passes through humanizeString (typo injection, casing tweaks). Mirrors the dialogue-history transform (stage 11).
• Uncensor input transforms — applyUncensorInputTransforms applies unicode-space injection and/or sanitization based on uncensorInputOptions.
• Mention conversion — every sample passes through convertMentions.
• Model sample prefix — ${botName}: ${output} prefix is prepended to the model side so the LLM learns to emit the bot-name prefix in its responses.

Invariants

After this stage runs:

• Returns [] if: impersonation, no tomoriState, empty arrays, or mismatched lengths between in/out.
• UNPAIRED_SAMPLE_DIALOGUE_SENTINEL on the user side skips the user item but still emits the model item (for “responses without a clear prompt” examples).
• Items are tagged DIALOGUE_SAMPLE so preset reassembly slots them before the real dialogue history (after which the LLM should “pivot” from sample voice to live conversation).

Configuration

Source	Field	Effect
tomoriState	sample_dialogues_in, sample_dialogues_out	The actual sample arrays
tomoriConfig	humanizer_degree	When HEAVY, applies humanizer to samples
tomoriConfig	uncensor_unicode_space_enabled, uncensor_sanitize_enabled	(Set upstream in nativeBuilder) Drives uncensor input transforms

Extension points

Surface	Plugin-relevance
UNPAIRED_SAMPLE_DIALOGUE_SENTINEL	Defined in src/types/preset/presetExport.ts; coupled to SillyTavern preset import.
Bot-name prefix on model samples	Hardcoded ${botName}: — a plugin adding “no-prefix mode” would extend here. → plugin plan candidate.
Humanizer + uncensor integration	Inherited from chat pipeline config; if a plugin adds new sample-transform passes (e.g. style transfer), it would extend the per-item transform sequence here.

This is the simplest contributor in the pipeline — it’s a pure formatter over two arrays. Most plugin-relevant work happens upstream in how sample_dialogues_in/out get populated (via persona editor, import, ST card conversion).

Related docs

• Persona sample editing: → no dedicated doc; persona configuration command surface only
• Humanizer system: → folded into dialogueHistory (stage 11); humanizeString helper in src/utils/text/processors/formatters.ts
• Uncensor system: → no dedicated doc; src/utils/text/uncensor.ts helper only
• ST preset sample-dialogue handling: docs/integrations/sillytavern/preset-system.md