LiteParse

lit extracts text from documents locally without a model. Every lit parse re-extracts the whole document, and every line you print stays in context for the rest of the session, so the goal is few parses, few turns, small output.

Parse once to a file, then search the file

lit parse "/abs/path/doc.pdf" --format text --no-ocr -o /tmp/doc.txt && wc -l /tmp/doc.txt

Reuse /tmp/doc.txt for every later question about the document. Never re-run lit parse to search.

• Born-digital PDF (has a text layer, like most reports): pass --no-ocr. It is much faster and gives the same text.
• Scanned PDF or image: drop --no-ocr. If digits look wrong, read the page visually (below).
• --format json only when you need bounding boxes or layout. Search it, never print it whole.

Search in few turns

• Get the context in the same command: grep -n -i -C4 "total assets" /tmp/doc.txt | head -40. Use sed -n 'A,Bp' only when you need a wider window than -C gave.

• Batch independent lookups into one command:

  for q in "carbon intensity" "scope 1" "total revenue"; do \
    echo "=== $q ==="; grep -n -i -C3 "$q" /tmp/doc.txt | head -25; done

• Bound every output with head. Use shell grep/sed on the saved file rather than the Read and Grep tools.

• If two targeted greps do not find the answer, run the bundled BM25 ranker once instead of trying keyword variants. Give it a natural-language query with several synonyms. -k is the number of matches and -e the lines of context around each:

  uv run --script <skill-dir>/scripts/search.py /tmp/doc.txt -q "materiality assessment priority topics" -k 8 -e 5

Read a page visually: last resort

Screenshots are the most expensive thing you can put in context. Use one only when text cannot answer the question (dense multi-column tables, figures, charts). Render one page at modest DPI, and re-render higher only if the text is illegible:

lit screenshot "/abs/path/doc.pdf" --target-pages "13" --dpi 150 -o /tmp/shots/   # then read the PNG

The page flag is --target-pages, not --pages.

Setup

Go straight to the parse. Skip lit --version and --help unless something failed. If lit is missing: npm i -g @llamaindex/liteparse. Office files need LibreOffice and images need ImageMagick; lit converts both to PDF.