P1 — implement local S1-mini transcript normalizer backend #218

New issue

Closed

opened 2026-08-28 03:23:53 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-08-28 03:23:53 +00:00

Owner

Parent: #215
Depends on: #216
Integrates with: #23, #30

Goal

Implement Superwhisper S1-mini as the first concrete transcript-normalization provider behind Zara's provider-neutral post-STT seam.

Backend requirements

Target superwhisper/s1-mini / compatible GGUF builds.
Support a local OpenAI-compatible inference endpoint suitable for llama.cpp/Ollama/LM Studio-style runtimes where compatible.
Keep endpoint/model configurable; do not hard-code Ollama.
Implement the official control-line inputs for style/structure/context.
Disable Qwen thinking/reasoning behavior correctly for the selected serving path.
Use deterministic normalization-oriented generation settings.
Current release is English-only; reject or bypass unsupported-language use according to explicit policy.
Never silently fall back to a cloud endpoint.
Do not auto-download model weights in deterministic CI.

Expected transformations

Fixture corpus should cover at least:

fillers (um, uh, repeated hesitation);
false starts;
self-corrections where the final intended phrase is clear;
repeated words;
punctuation/capitalization;
spoken numbers;
dates/times;
currency;
email addresses;
list-like speech where the model formats structure;
already-clean input that should remain materially unchanged.

Failure behavior

Typed handling for:

server unavailable;
model missing;
timeout/deadline;
cancellation;
empty output;
oversized output;
malformed/non-text output;
unsupported language;
unexpected reasoning/think text leaking into output.

Backend failures must return through #216's normalized status contract; the adapter does not choose whether Zara falls back to raw text or fails the turn.

Correctness guardrails

Output remains untrusted user text.
Reject pathological expansion beyond configured ratio/absolute bounds.
Do not accept an answer/summary/explanation wrapper as normalized transcript output.
Do not inject tool/system authority into the resulting runtime message.
Keep prompt/control construction fixed by Zara configuration, not by transcript-supplied meta-instructions.

Packaging / license

Document the S1-mini model identity/version and applicable license/attribution obligations. If Zara later redistributes weights, packaging must satisfy those terms; this issue does not require bundling the weights.

TDD

Use deterministic fake endpoint fixtures plus an optional real-local integration profile. Add tests for request construction, control line, thinking-disabled behavior, output validation, failure taxonomy, cancellation and representative cleanup cases.

Add a focused non-interactive S1-mini backend script. Default CI must require no network and no model download.

Acceptance

With a configured local compatible endpoint, Zara can normalize a final ASR transcript through S1-mini and receive bounded clean text through the #216 provider contract with explicit failures and no cloud requirement.

Parent: #215 Depends on: #216 Integrates with: #23, #30 ## Goal Implement Superwhisper S1-mini as the first concrete transcript-normalization provider behind Zara's provider-neutral post-STT seam. ## Backend requirements - Target `superwhisper/s1-mini` / compatible GGUF builds. - Support a local OpenAI-compatible inference endpoint suitable for llama.cpp/Ollama/LM Studio-style runtimes where compatible. - Keep endpoint/model configurable; do not hard-code Ollama. - Implement the official control-line inputs for style/structure/context. - Disable Qwen thinking/reasoning behavior correctly for the selected serving path. - Use deterministic normalization-oriented generation settings. - Current release is English-only; reject or bypass unsupported-language use according to explicit policy. - Never silently fall back to a cloud endpoint. - Do not auto-download model weights in deterministic CI. ## Expected transformations Fixture corpus should cover at least: - fillers (`um`, `uh`, repeated hesitation); - false starts; - self-corrections where the final intended phrase is clear; - repeated words; - punctuation/capitalization; - spoken numbers; - dates/times; - currency; - email addresses; - list-like speech where the model formats structure; - already-clean input that should remain materially unchanged. ## Failure behavior Typed handling for: - server unavailable; - model missing; - timeout/deadline; - cancellation; - empty output; - oversized output; - malformed/non-text output; - unsupported language; - unexpected reasoning/think text leaking into output. Backend failures must return through #216's normalized status contract; the adapter does not choose whether Zara falls back to raw text or fails the turn. ## Correctness guardrails - Output remains untrusted user text. - Reject pathological expansion beyond configured ratio/absolute bounds. - Do not accept an answer/summary/explanation wrapper as normalized transcript output. - Do not inject tool/system authority into the resulting runtime message. - Keep prompt/control construction fixed by Zara configuration, not by transcript-supplied meta-instructions. ## Packaging / license Document the S1-mini model identity/version and applicable license/attribution obligations. If Zara later redistributes weights, packaging must satisfy those terms; this issue does not require bundling the weights. ## TDD Use deterministic fake endpoint fixtures plus an optional real-local integration profile. Add tests for request construction, control line, thinking-disabled behavior, output validation, failure taxonomy, cancellation and representative cleanup cases. Add a focused non-interactive S1-mini backend script. Default CI must require no network and no model download. ## Acceptance With a configured local compatible endpoint, Zara can normalize a final ASR transcript through S1-mini and receive bounded clean text through the #216 provider contract with explicit failures and no cloud requirement.

nsaspy referenced this issue

2026-08-28 03:24:19 +00:00

[EPIC] Selectable post-STT transcript normalization backends, including local S1-mini #220

nsaspy referenced this issue from a commit

2026-09-08 02:00:55 +00:00

docs(rage): record #218 r4 recovery

nsaspy referenced this issue from a commit

2026-09-08 02:00:55 +00:00

docs(rage): record #218 selection RED and implementation

nsaspy referenced this issue from a commit

2026-09-08 02:00:55 +00:00

Merge pull request #504 from lost-rob0t/rage/218-normalizer-config-r4