wdoc
Written for wdoc v5.1.0. On a different version, some arguments, defaults, or behaviors may differ.
wdoc is a RAG (Retrieval-Augmented Generation) system for summarizing, searching, and querying documents across 20+ file types. It works as a CLI (via Google Fire) and as a Python library (from wdoc import wdoc), routing every LLM call through LiteLLM (100+ providers).
This SKILL.md is the quick orientation. The deep material lives in two companion files:
- REFERENCE.md: every CLI argument, filetype, loader option, environment variable, and the full Python API.
- EXAMPLES.md: copy-pasteable shell and Python examples.
Quick start
bashpip install -U wdoc[full] # full install: all loaders. Plain `wdoc` ships only PDF + URL. export ANTHROPIC_API_KEY="your_key" # or whichever provider you use wdoc query paper.pdf "What are the main findings?" # ask questions (RAG) wdoc summarize paper.pdf # detailed markdown summary wdoc parse paper.pdf # parse to text, no LLM wdoc web "latest on quantum computing" # DuckDuckGo + query
uvx wdoc[full] ... runs it without installing and sidesteps thinking about extras.
The four tasks
| Task | What it does | Pick it when |
|---|---|---|
query | Embeds docs, retrieves chunks, answers with sourced markdown | You have a question about the content |
search | Returns matching docs + metadata, no LLM answer | You only need to locate relevant passages |
summarize | Detailed markdown summary (author's reasoning, not vague takeaways) | You want the gist of a long document |
summarize_then_query | Summarize first, then drop into a query prompt | You want both, in one run |
Core mechanics worth knowing
- Shortcuts:
wdoc query FILE,wdoc summarize FILE,wdoc parse FILE, andwdoc web "q"expand to longer--task=...forms. Positional args work too:wdoc TASK PATH [QUERY]. - Filetype is auto-detected (
--filetype=auto) but can be forced (pdf,youtube,anki,zotero, ...). Recursive filetypes (recursive_paths,zotero,karakeep,ddg, ...) fan one selector out into many documents. - Two models per run: a strong
--modelanswers, a cheap--query_eval_modelfilters chunks. Both take LiteLLMprovider/modelids. - kebab or snake case:
--query-eval-modeland--query_eval_modelare equivalent. - Piped input is auto-detected:
cat file.pdf | wdoc parse --filetype=pdf. - Privacy:
--private(orWDOC_PRIVATE_MODE=true) blocks all outbound traffic and redacts API keys; pair it with local models (Ollama) and--llms_api_bases. - Reuse embeddings:
--save_embeds_as=idx.pklonce, then--load_embeds_from=idx.pklto skip re-indexing. - Cost guard:
--dollar_limit(default 5) stops summaries/embeddings before they get expensive.
Common patterns
bash# Query every PDF in a tree wdoc --task=query --path="papers/" --filetype=recursive_paths \ --pattern="**/*.pdf" --recursed_filetype=pdf --query="..." # Fully local / private wdoc --private --model="ollama/qwen3:8b" --query_eval_model="ollama/qwen3:8b" \ --embed_model="ollama/snowflake-arctic-embed2" --task=query --path=secret.pdf # Parse for use elsewhere (text, langchain, langchain_dict, xml, split_text) wdoc parse document.pdf --format=langchain_dict
pythonfrom wdoc import wdoc instance = wdoc(task="query", path="paper.pdf", model="openai/gpt-4o") answer = instance.query_task("What are the main contributions?") print(answer["final_answer"])
For anything beyond this page (exact argument types, defaults, every filetype's loader options, all WDOC_* env vars, the full Python API surface, and more examples), read REFERENCE.md and EXAMPLES.md.
