Point it at your own LLM (local Ollama / vLLM / LM Studio / DGX Spark all work),
and these 13 tools gain smart options. The core tools do not depend on an LLM at all; without any of this they still work 100%.
Sentence translation
Sentences are sent to the LLM, source and translation side by side, each re-translatable on its own. A Taiwanese IT terminology list is built in, and an optional “document domain” hint (legal / medical / technical) sharpens the vocabulary.
Result: for long contracts and papers, translating sentences in parallel beats doing it by hand
Document translation (whole office documents)
Translates a whole Word, Excel or PowerPoint file into another language and produces a file in the same format with the same layout: only the text changes; borders, tables, headers, footers and images all stay where they are. Paragraphs are sent to the LLM as units, and the first 6 pages come back as a side-by-side preview.
Result: a translated contract or policy document straight out, with no re-layout needed
Extract text (paragraph re-flow)
After PyMuPDF extracts the text, each block goes to the LLM to rejoin sentences that line breaks had split. Exports TXT / Markdown / DOCX / ODT.
Result: body text chopped up by the PDF layout reads smoothly again
Auto-fill forms (verification)
After filling, the LLM reviews the rendered result (PNG) field by field, spotting misplaced or truncated values and wrongly ticked checkboxes.
Result: a 30-column vendor form needs no column-by-column proofreading; just look at the ones the LLM flagged
Document redaction (extra detection)
Regex catches fixed formats (ID numbers, phone numbers, bank accounts, tax IDs); the LLM adds context-sensitive cases such as “customer code A-2024-0815” or “Manager Wang”.
Result: the LLM catches the non-standard formats and context-dependent fields that a regex cannot
Text redaction (extra detection)
The same, but for plain text: support conversations, logs, email bodies. Paste it in or upload.txt /.md.
Result: plain text gets the same regex + LLM double detection
Word count (summary / keywords)
After counting, the LLM can add a summary and a list of key concepts (optional, off by default).
Result: faced with an unfamiliar long document, read the LLM summary first and then decide whether to read it properly
Meeting summary
Turns a meeting transcript into a summary, decisions, action items, risks and chapters. Every entry carries a segment number; click it to jump back to that line. Minutes get used as the record of what was agreed, so a decision with no source is worse than no decision at all.
Annotation report (automatic grouping)
Every annotation in the PDF is pulled out and the LLM groups them by theme (text to change, formatting issues, open questions, already agreed … it decides freely).
Result: on a long document reviewed by several people, group similar suggestions in one click
Document compare (summary of changes)
After the line-by-line diff, the LLM adds three to five sentences explaining what changed overall (optional, off by default).
Result: after a contract, draft or specification is revised, read the LLM summary before deciding whether to go through it clause by clause
OCR (typo correction)
After EasyOCR or Tesseract runs, the LLM uses the surrounding context to fix common OCR mistakes. A hallucination guard is built in: corrections apply only when the word count is identical before and after, so the LLM cannot rewrite freely.
Result: cleaner OCR output from scans, ready to search or edit
Pre-submission check (meaning + visual)
After the structural checks (page size, embedded fonts), the LLM checks the content for meaning (are required fields really filled, is the wording self-consistent) and a vision model inspects the PNG (does a stamp cover text, is the layout broken).
Result: catches the "right characters, wrong meaning" and "looks wrong" cases that regex and structural checks miss
e-Invoice processing (accounting category)
After scanning, invoices are sent to the LLM in batches: by seller tax ID, company name and industry: to decide the accounting category (fuel, meals, postage …). Rules run first and the LLM fills the gaps; your own rules can override both.
Result: classified automatically before filing an expense claim, with no picking of accounts by hand
