docqa

name	docqa
description	Parse large PDF documents and answer questions with accurate citations from the PDF content.

DocQA

Parse large PDF documents, such as technical manuals, and answer questions by accurately citing document content.

Preflight

Before parsing:

1. Ensure current system is Linux or macOS.
2. Ensure Deno is installed. If not, stop and ask for permission to install with curl -fsSL https://deno.land/install.sh | sh.
3. Ensure PDF-related tools are installed. If not, stop and ask for permission to install poppler.
4. Check whether ~/.agents/skills/docqa/.env is configured. It must contain the OpenAI-compatible endpoint configuration for the image-processing model:
   • OPENAI_BASE_URL
   • OPENAI_API_KEY
   • OPENAI_MODEL
5. If ~/.agents/skills/docqa/.env does not exist or any required value is missing, ask the user for the OpenAI base URL, API key, and image-processing model name, then write them to ~/.agents/skills/docqa/.env.

Index a PDF

When the user asks to process or index a PDF document:

1. Resolve the PDF path the user wants to process.

2. Run the CLI index command:

   deno run --allow-read --allow-write --allow-env --allow-run --allow-sys --allow-net ~/.agents/skills/docqa/scripts/cli <pdf-file> index
   

3. Wait for the command to finish. On success, report the full cache directory path printed by the CLI.

Answer Questions

When the user asks a question about an indexed PDF:

1. Use the cache directory produced by the indexing step.
2. Search index/ with jq to inspect page-level keywords, summaries, topics, headings, facts, and other structured index fields. Use the user's question to identify multiple pages that are most likely relevant.
3. After selecting candidate pages, read the corresponding original extracted Markdown files from markdown/, such as markdown/page-0001.md.
4. Answer the user's question faithfully from the Markdown page content. Do not rely only on the index summaries, and do not invent facts not supported by the extracted page text.
5. Include the supporting PDF page numbers in the answer, using the page numbers from the index or Markdown file names.