Skip to content

Filter LLM Output Artifacts

Strip non-spoken roleplay markup from streamed Generate / GenerateAndSpeak output before it reaches the client, dialogue history, or TTS: speaker-name prefixes (Barnaby: …), asterisk emotes (*sighs*), and parenthetical / bracket stage directions.

All flags default off. Filtering is opt-in per node. Transformed text is authoritative for the slot, wire answer, history, and TTS.

Prerequisites

  • A Generate or GenerateAndSpeak node.
  • Willingness to mutate history: filtered text is what the agent remembers saying on later turns.

1. Enable the filter on the node

Set fields under output_filter (OutputFilterOverrides):

Field Default Purpose
strip_speaker_prefix false Remove a leading Name: tag once per turn
max_prefix_bytes 16 UTF-8 byte budget for the prefix scan ([1, 64])
strip_asterisk_spans false Delete *…* spans (first-close wins)
strip_paren_spans false Delete (…) spans
strip_bracket_spans false Delete […] spans
max_span_bytes 64 UTF-8 byte budget while matching a span ([1, 256]) 
from tryll_client.graph import GenerateParams, OutputFilterOverrides, SendAnswer

GenerateParams(
    send=SendAnswer.Streamed,
    output_filter=OutputFilterOverrides(
        strip_speaker_prefix=True,
        strip_asterisk_spans=True,
        strip_paren_spans=True,
    ),
)

On the Generate / GenerateAndSpeak node params, expand Output Filter and enable the flags you need. Budgets are clamped in the inspector.

2. What the rules do

  • Speaker prefix (R1). At the first non-whitespace byte of the turn, if a : followed by whitespace appears within the budget, and the candidate starts with A–Z, has at most three ASCII words, and contains no . ! ? ,, the tag is deleted. ASCII-only. Residual false positives such as One rule: … are possible — keep the flag off unless you want this cleanup.
  • Paired spans (R2). Enabled delimiter pairs delete open…close including content. First close wins; nested opens are treated as content. Unmatched or over-budget candidates are emitted unchanged (benign failure).
  • Whitespace. Leading/trailing ASCII whitespace is trimmed; a space on both sides of a deleted span squeezes to one space when still stream-safe.

Filtering runs before substitute_agent_variables. Markers inside a deleted span are never substituted; values introduced by substitution are never re-filtered.

3. Diagnostics

When enable_diagnostics is on, each Generate / GenerateAndSpeak node entry may include output_filter_removals:

{
  "rule": "strip_speaker_prefix",
  "matched": "Barnaby: ",
  "source_byte_offset": 0,
  "output_byte_offset": 0
}

At most 256 records are kept (output_filter_removals_truncated when exceeded). Telemetry omits matched content when user-content telemetry is off.

When not to use this

  • Leave flags off for free-form chat that may legitimately start with Label: value lines.
  • Do not expect curly quotes, Unicode names, or third-person narration without markup to be cleaned — those are prompt/eval concerns or future flags.
  • Quote wrapping ("…" around a whole turn) is not stripped in this release.

See also Substitute Agent Variables in LLM Output for the sibling stream transform.