scriptflow/docs/plans/2026-03-11-upload-extract-episodes-design.md

Upload And Extract Episodes Design

Context

The conversion mode currently accepts only manual text input in the left textarea. The app already has a Doubao streaming integration pattern for script generation, and the extracted content should feed back into the existing sourceText flow rather than replacing the rest of the conversion pipeline.

Decision

Adopt a client-side upload flow for four file types: Word (.docx), text (.txt), PDF (.pdf), and Markdown (.md). After upload, the app will read the file in the browser, send the raw text to a new Doubao extraction call using doubao-seed-1-6-flash-250828, and stream the model output directly into the left-side source textarea.

Behavior

• The source input area becomes a hybrid input surface: manual typing still works, and file upload is added alongside it.
• Upload immediately starts extraction without requiring the user to click 立即转换成剧本.
• The extraction model is instructed to identify each episode and return the original script content 1:1 with no rewriting, normalization, cleanup, or omission.
• The streamed extraction result overwrites sourceText progressively so the user can see the result arrive in real time.
• Existing conversion generation stays separate. After extraction completes, the user can still click the existing conversion button to continue with the current workflow.

Parsing Strategy

• .txt and .md: read with File.text().
• .docx: parse in-browser with a document-text extraction library.
• .pdf: parse in-browser with a PDF text extraction library.

UI And State

• Add upload affordance, accepted-file hint, extraction loading state, and extraction error state in conversion mode.
• Preserve sourceText local persistence.
• Keep manual editing enabled after extraction.

AI Contract

The new extraction API will:

• use Doubao only
• stream results
• instruct the model to output episode-separated original content only
• avoid any transformations beyond episode boundary recognition

Risks

• PDF text extraction quality depends on document structure.
• Even with strict prompting, model-based extraction is probabilistic, so the prompt must strongly prohibit edits and define a deterministic output format.
• Browser-side parsing adds dependency and bundle-size cost.