Document extraction
The development scope includes HWP/HWPX, PDF, Office files, images, text, and email. OCR is used for supported image-based material.
FINDOC
A desktop document search app in development for local folders and NAS shares, for people who need to locate conditions, dates, and clauses across HWP, PDF, and Office files.
In development · Preparing for release
Local document engine
The desktop workflow keeps extraction, indexing, and search on the device. The planned Claude API feature is separate and will only use passages a user explicitly selects.
Findoc is designed around folders, shared drives, scanned papers, and office files people already use.
The development scope includes HWP/HWPX, PDF, Office files, images, text, and email. OCR is used for supported image-based material.
After indexing, keyword and semantic search can run on the user’s device. A network connection is not required for local searching.
File type, date, and path filters help users move from a broad result list to a document they can inspect.
Recorded checks cover document extraction; full search-workflow evaluation comes next.
A Codex run produced non-empty extraction for one fixture each of PDF, DOCX, HWPX, Markdown, TXT, HTML, JSON, and EML. It also found the expected keywords in the TXT and Markdown fixtures, and 12 selected parser tests passed.
This does not validate the full GUI search flow, search accuracy, or real user use. Those need separate evaluation with permitted samples.
We plan to run one Findoc document-to-search task through requirements, implementation, independent QA, and founder approval. With permitted samples, we will assess search accuracy, task completion, and cost before reusing the workflow for other products and custom software projects.
The path from file to result stays visible.
Select local folders or a configured shared location.
Extract available text and prepare it for search.
Review results, filters, and the original document.
Illustrative planned workflow
Question: “When is the renewal notice due?”
Answer: cited clause from a selected contract, or withheld if unsupported
Next: open the original file to verify context
Findoc is currently developed with Codex. We are evaluating a move to Claude Code for selected development and verification work; a Claude API feature remains a separate, opt-in cloud query flow.
Codex supports requirement structuring, implementation and documentation review, and verification steps. Claude Code is under evaluation for selected workflows. Release decisions remain with the founder.
A user would search relevant passages locally, see the proposed transmission range, and consent before a server sends a question and the minimum selected text to the Claude API. The app would reconfirm the scope before sending, apply a request-cost cap, and keep local search available if the request fails.
The planned response will cite a file and passage, reducing time spent comparing several files while keeping the original source one click away. When selected material does not support an answer, it should withhold an answer.
Local search is designed to work offline. Any future cloud question flow is separate, shows its scope, and requires a user choice before sending selected text.
Before expansion, we plan to assess source accuracy, appropriate “I don’t know” responses, latency, and request cost. Other Claude API ideas will wait for this evaluation. No performance level is promised here.
Product conversation