Skip to content

Latest commit

 

History

History
332 lines (292 loc) · 30 KB

README.md

File metadata and controls

332 lines (292 loc) · 30 KB
Raw

I'm wdoc. I solve RAG problems.

  • wdoc, imitating Winston "The Wolf" Wolf

wdoc

wdoc is a powerful RAG (Retrieval-Augmented Generation) system designed to summarize, search, and query documents across various file types. It's particularly useful for handling large volumes of diverse document types, making it ideal for researchers, students, and professionals dealing with extensive information sources. I was frustrated with all other RAG solutions for querying or summarizing, so I made my perfect solution in a single package.

  • Goal and project specifications: wdoc uses LangChain to process and analyze documents. It's capable of querying tens of thousands of documents across various file types at the same time. The project also includes a tailored summary feature to help users efficiently keep up with large amounts of information.

  • Current status: Under active development

    • Used daily by the developer for several months: but still in alpha. I would greatly benefit from testing by users as it's the quickest way for me to find the many small bugs that are quick to fix.
    • May have some instabilities, but issues can usually be resolved quickly
    • The main branch is more stable than the dev branch, which offers more features
    • Open to feature requests and pull requests
    • All feedbacks, including reports of typos, are highly appreciated
    • Please open an issue before making a PR, as there may be ongoing improvements in the pipeline

  • Key Features:

    • Aims to support any filetypes and query from all of them at the same time (15+ are already implemented!)
    • High recall and specificity: it was made to find A LOT of documents using carefully designed embedding search then carefully aggregate gradually each answer using semantic batch to produce a single answer that mentions the source poiting to the exact portion of the source document.
    • Supports virtually any LLM, including local ones, and even with extra layers of security for super secret stuff.
    • Use both an expensive and cheap LLM to make recall as high as possible because we can afford fetching a lot of documents per query (via embeddings)
    • At last a usable text summary: get the thought process of the author instead of nebulous takeaways.
    • Extensible, this is both a tool and a library.

Table of contents

Ultra short guide for people in a hurry

Give it to me I am in a hurry! 
link="https://situational-awareness.ai/wp-content/uploads/2024/06/situationalawareness.pdf"

wdoc --path=$link --task=query --filetype="online_pdf" --query="What does it say about alphago?" --query_retrievers='default_multiquery' --top_k=auto_200_500
  • This will:
    1. parse what's in --path as a link to a pdf to download (otherwise the url could simply be a webpage, but in most cases you can leave it to 'auto' by default as heuristics are in place to detect the most appropriate parser).
    2. cut the text into chunks and create embeddings for each
    3. Take the user query, create embeddings for it ('default') AND ask the default LLM to generate alternative queries and embed those
    4. Use those embeddings to search through all chunks of the text and get the 200 most appropriate documents
    5. Pass each of those documents to the smaller LLM (default: openai/gpt-4o-mini) to tell us if the document seems appropriate given the user query
    6. If More than 90% of the 200 documents are appropriate, then we do another search with a higher top_k and repeat until documents start to be irrelevant OR we it 500 documents.
    7. Then each relevant doc is sent to the strong LLM (by default, openai/gpt-4o) to extract relevant info and give one answer.
    8. Then all those "intermediate" answers are 'semantic batched' (meaning we create embeddings, do hierarchical clustering, then create small batch containing several intermediate answers) and each batch is combined into a single answer.
    9. Rinse and repeat steps 7+8 until we have only one answer, that is returned to the user.
link="https://situational-awareness.ai/wp-content/uploads/2024/06/situationalawareness.pdf"

wdoc --path=$link --task=summarize --filetype="online_pdf"

  • This will:

    1. Split the text into chunks
    2. pass each chunk into the strong LLM (by default openai/gpt-4o) for a very low level (=with all details) summary. The format is markdown bullet points for each idea and with logical indentation.
    3. When creating each new chunk, the LLM has access to the previous chunk for context.
    4. All summary are then concatenated and returned to the user

  • For extra large documents like books for example, this summary can be recusively fed to wdoc using argument --summary_n_recursion=2 for example.

  • Those two tasks, query and summary, can be combined with --task summarize_then_query which will summarize the document but give you a prompt at the end to ask question in case you want to clarify things.

  • For more, you can jump to the section Walkthrough and examples

Features

  • 15+ filetypes: also supports combination to load recursively or define complex heterogenous corpus like a list of files, list of links, using regex, youtube playlists etc. See Supported filestypes. All filetype can be seamlessly combined in the same index, meaning you can query your anki collection at the same time as your work PDFs). It supports removing silence from audio files and youtube videos too!
  • 100+ LLMs: OpenAI, Mistral, Claude, Ollama, Openrouter, etc. Thanks to litellm. Personnaly I'm using openrouter's Sonnet 3.5 as strong LLM and openai's gpt-4o-mini as query_eval LLM, with openai embeddings.
  • Local and Private LLM: take some measures to make sure no data leaves your computer and goes to an LLM provider: no API keys are used, all api_base are user set, cache are isolated from the rest, outgoing connections are censored by overloading sockets, etc.
  • Advanced RAG to query lots of diverse documents:
    1. The documents are retrieved using embedding
    2. Then a weak LLM model ("Eve the Evaluator") is used to tell which of those document is not relevant
    3. Then the strong LLM is used to answer ("Anna the Answerer") the question using each individual remaining documents.
    4. Then all relevant answers are combined ("Carl the Combiner") into a single short markdown-formatted answer. Before being combined, they are batched by semantic clusters and semantic order using scipy's hierarchical clustering and leaf ordering, this makes it easier for the LLM to combine the answers in a manner that makes bottom up sense. Eve the Evaluator, Anna the Answerer and Carl the Combiner are the names given to each LLM in their system prompt, this way you can easily add specific additional instructions to a specific step. There's also Sam the Summarizer for summaries and Raphael the Rephraser to expand your query.
    5. Each document is identified by a unique hash and the answers are sourced, meaning you know from which document comes each information of the answer.
    • Supports a special syntax like "QE >>>> QA" were QE is a question used to filter the embeddings and QA is the actual question you want answered.
  • Advanced summary:
    • Instead of unusable "high level takeaway" points, compress the reasoning, arguments, though process etc of the author into an easy to skim markdown file.
    • The summaries are then checked again n times for correct logical indentation etc.
    • The summary can be in the same language as the documents or directly translated.
  • Many tasks: See Supported tasks.
  • Trust but verify: The answer is sourced: wdoc keeps track of the hash of each document used in the answer, allowing you to verify each assertion.
  • Markdown formatted answers and summaries: using rich.
  • Sane embeddings: By default use sophisticated embeddings like multi query retrievers but also include SVM, KNN, parent retriever etc. Customizable.
  • Fully documented Lots of docstrings, lots of in code comments, detailed --help etc. The full usage can be found in the file USAGE.md or via python -m wdoc --help. I work hard to maintain an exhaustive documentation.
  • Scriptable / Extensible: You can use wdoc in other python project using --import_mode. Take a look at the scripts below.
  • Statically typed: Runtime type checking. Opt out with an environment flag: WDOC_TYPECHECKING="disabled / warn / crash" wdoc (by default: warn). Thanks to beartype it shouldn't even slow down the code!
  • LLM (and embeddings) caching: speed things up, as well as index storing and loading (handy for large collections).
  • Good PDF parsing PDF parsers are notoriously unreliable, so 15 (!) different loaders are used, and the best according to a parsing scorer is kept. Including table support via openparse (no GPU needed by default) or via UnstructuredPDFLoader.
  • Langfuse support: If you set the appropriate langfuse environment variables they will be used. See this guide or this one to learn more (Note: this is disabled if using private_mode to avoid any leaks).
  • Document filtering: based on regex for document content or metadata.
  • Fast: Parallel document loading, parsing, embeddings, querying, etc.
  • Shell autocompletion using python-fire
  • Notification callback: Can be used for example to get summaries on your phone using ntfy.sh.
  • Hacker mindset: I'm a friendly dev! Just open an issue if you have a feature request or anything else.

Planned features

Click to read more

(These don't include improvements, bugfixes, refactoring etc.)

  • THIS LIST IS NOT UP TO DATE AND THERE ARE MANY MORE THINGS PLANNED
  • Start using unit tests
  • Accept input from stdin, to for example query directly from a manpage
  • Much faster startup time
  • Much improved retriever:
    • Web search retriever, online information lookup via jina.ai reader and search.
    • LLM powered synonym expansion for embeddings search.
  • A way to specify at indexing time how trusting you are of a given set of document.
  • A way to open the documents automatically, based on the platform used. For ex if okular is installed, open pdfs directly at the appropriate page.
  • Improve the scriptability of wdoc. Add examples for how you use it with Logseq.
    • Include a server example, that mimics the OpenAI's API to make your RAG directly accessible to other apps.
    • Add a gradio GUI.
  • Include the possible whisper/deepgram extra expenses when counting costs.
  • Add support for user defined loaders.
  • Automatically caption document images using an LLM, especially nice for anki cards.

Supported filetypes

  • auto: default, guess the filetype for you

  • url: try many ways to load a webpage, with heuristics to find the better parsed one

  • youtube: text is then either from the yt subtitles / translation or even better: using whisper / deepgram. Note that youtube subtitles are downloaded with the timecode (so you can ask 'when does the author talks about such and such) but at a lower sampling frequency (instead of one timecode per second, only one per 15s). Youtube chapters are also given as context to the LLM when summarizing, which probably help it a lot.

  • pdf: 15 default loaders are implemented, heuristics are used to keep the best one and stop early. Table support via openparse or UnstructuredPDFLoader. Easy to add more.

  • online_pdf: via URL then treated as a pdf (see above)

  • anki: any subset of an anki collection db. alt and title of images can be shown to the LLM, meaning that if you used the ankiOCR addon this information will help contextualize the note for the LLM.

  • string: the cli prompts you for a text so you can easily paste something, handy for paywalled articles!

  • txt: .txt, markdown, etc

  • text: send a text content directly as path

  • local_html: useful for website dumps

  • logseq_markdown: thanks to my other project: LogseqMarkdownParser you can use your Logseq graph

  • local_audio: supports many file formats, can use either OpenAI's whisper or deepgram. Supports automatically removing silence etc.

  • local_video: extract the audio then treat it as local_audio

  • online_media: use youtube_dl to try to download videos/audio, if fails try to intercept good url candidates using playwright to load the page. Then processed as local_audio (but works with video too).

  • epub: barely tested because epub is in general a poorly defined format

  • powerpoint: .ppt, .pptx, .odp, ...

  • word: .doc, .docx, .odt, ...

  • json_dict: a text file containing a single json dict.

  • Recursive types

    • youtube playlists: get the link for each video then process as youtube
    • recursive_paths: turns a path, a regex pattern and a filetype into all the files found recurisvely, and treated a the specified filetype (for example many PDFs or lots of HTML files etc).
    • link_file: turn a text file where each line contains a url into appropriate loader arguments. Supports any link, so for example webpage, link to pdfs and youtube links can be in the same file. Handy for summarizing lots of things!
    • json_entries: turns a path to a file where each line is a json dict: that contains arguments to use when loading. Example: load several other recursive types. An example can be found in docs/json_entries_example.json.
    • toml_entries: read a .toml file. An example can be found in docs/toml_entries_example.toml.

Supported tasks

  • query give documents and asks questions about it.
  • search only returns the documents and their metadata. For anki it can be used to directly open cards in the browser.
  • summarize give documents and read a summary. The summary prompt can be found in utils/prompts.py.
  • summarize_then_query summarize the document then allow you to query directly about it.

Walkthrough and examples

Detailed example
  1. Say you want to ask a question about one pdf, that's simple: wdoc --task="query" --path="my_file.pdf" --filetype="pdf" --modelname='openai/gpt-4o'. Note that you could have just let --filetype="auto" and it would have worked the same.
  • Note: By default wdoc tries to parse args as kwargs so wdoc query mydocument What's the age of the captain? is parsed as wdoc --task=query --path=mydocument --query "What's the age of the captain?". Likewise for summaries. This does not always work so use it only after getting comfortable with wdoc.
  1. You have several pdf? Say you want to ask a question about any pdf contained in a folder, that's not much more complicated : wdoc --task="query" --path="my/other_dir" --pattern="**/*pdf" --filetype="recursive_paths" --recursed_filetype="pdf" --query="My question about those documents". So basically you give as path the path to the dir, as pattern the globbing pattern used to find the files relative to the path, set as filetype "recursive_paths" so that wdoc knows what arguments to expect, and specify as recursed_filetype "pdf" so that wdoc knows that each found file must be treated as a pdf. You can use the same idea to glob any kind of file supported by wdoc like markdown etc. You can even use "auto"! Note that you can either directly ask your question with --query="my question", or wait for an interactive prompt to pop up, or just pass the question as *args like so wdoc [your kwargs] here is my question.
  2. You want more? You can write a .json file where each line (#comments and empty lines are ignored) will be parsed as a list of argument. For example one line could be : {"path": "my/other_dir", "pattern": "**/*pdf", "filetype": "recursive_paths", "recursed_filetype": "pdf"}. This way you can use a single json file to specify easily any number of sources. .toml files are also supported.
  3. You can specify a "source_tag" metadata to help distinguish between documents you imported. It is EXTREMELY recommended to include a source_tag to any document you want to save: especially if using recursive filetypes. This is because after loading all documents wdoc use the source_tag to see if it should continue or crash. If you want to load 10_000 pdf in one go as I do, then it makes sense to continue if some failed to crash but not if a whole source_tag is missing.
  4. Now say you do this with many many documents, as I do, you of course can't wait for the indexing to finish every time you have a question (even though the embeddings are cached). You should then add --save_embeds_as=your/saving/path to save all this index in a file. Then simply do --load_embeds_from=your/saving/path to quickly ask queries about it!
  5. To know more about each argument supported by each filetype, wdoc --help
  6. There is a specific recursive filetype I should mention: --filetype="link_file". Basically the file designated by --path should contain in each line (#comments and empty lines are ignored) one url, that will be parsed by wdoc. I made this so that I can quickly use the "share" button on android from my browser to a text file (so it just appends the url to the file), this file is synced via syncthing to my browser and wdoc automatically summarize them and add them to my Logseq. Note that the url is parsed in each line, so formatting is ignored, for example it works even in markdown bullet point list.
  7. If you want to make sure your data remains private here's an example with ollama: wdoc --private --llms_api_bases='{"model": "http://localhost:11434", "query_eval_model": "http://localhost:11434"}' --modelname="ollama_chat/gemma:2b" --query_eval_modelname="ollama_chat/gemma:2b" --embed_model="BAAI/bge-m3" my_task
  8. Now say you just want to summarize Tim Urban's TED talk on procrastination: wdoc --task=summary --path='https://www.youtube.com/watch?v=arj7oStGLkU' --youtube_language="english" --disable_md_printing:
Click to see the output

Summary

https://www.youtube.com/watch?v=arj7oStGLkU

  • The speaker, Tim Urban, was a government major in college who had to write many papers
  • He claims his typical work pattern for papers was:
    • Planning to spread work evenly
    • Actually procrastinating until the last minute
  • For his 90-page senior thesis:
    • Planned to work steadily over a year
    • Actually ended up writing 90 pages in 72 hours before the deadline
    • Pulled two all-nighters
    • Resulted in a 'very, very bad thesis'
  • Urban is now a writer-blogger for 'Wait But Why'
  • He wrote about procrastination to explain it to non-procrastinators
  • Humorously claims to have done brain scans comparing procrastinator and non-procrastinator brains
  • Introduces concept of 'Instant Gratification Monkey' in procrastinator's brain
    • Monkey takes control from the Rational Decision-Maker
    • Leads to unproductive activities like reading Wikipedia, checking fridge, YouTube spirals
  • Monkey characteristics:
    • Lives in the present moment
    • No memory of past or knowledge of future
    • Only cares about 'easy and fun'
  • Rational Decision-Maker:
    • Allows long-term planning and big picture thinking
    • Wants to do what makes sense in the moment
  • 'Dark Playground': where procrastinators spend time on leisure activities when they shouldn't
    • Filled with guilt, dread, anxiety, self-hatred
  • 'Panic Monster': procrastinator's guardian angel
    • Wakes up when deadlines are close or there's danger of embarrassment
    • Only thing the Monkey fears
  • Urban relates his own experience procrastinating on preparing this TED talk
  • Claims thousands of people emailed him about having the same procrastination problem
  • Two types of procrastination:
      1. Short-term with deadlines (contained by Panic Monster)
      1. Long-term without deadlines (more damaging)
      • Affects self-starter careers, personal life, health, relationships
      • Can lead to long-term unhappiness and regrets
  • Urban believes all people are procrastinators to some degree
  • Presents 'Life Calendar': visual representation of weeks in a 90-year life
  • Encourages audience to:
    • Think about what they're procrastinating on
    • Stay aware of the Instant Gratification Monkey
    • Start addressing procrastination soon
  • Humorously suggests not starting today, but 'sometime soon'

Tokens used for https://www.youtube.com/watch?v=arj7oStGLkU: '4365' ($0.00060)

Total cost of those summaries: '4365' ($0.00060, estimate was $0.00028)

Total time saved by those summaries: 8.4 minutes

Done summarizing.

Getting started

wdoc currently requires Python 3.11 to run for now. Make sure that your Python version matches this one or it will not work.

  1. To install:
    • Using pip: pip11 install -U wdoc
    • Or to get a specific git branch:
      • dev branch: pip11 install git+https://github.com/thiswillbeyourgithub/wdoc@dev
      • main branch: pip11 install git+https://github.com/thiswillbeyourgithub/wdoc@main
    • You can also use pipx or uvx. But as I'm not experiences with them I don't know if that can cause issues with for example caching etc. Do tell me if you tested it!
      • Using pipx: pipx run wdoc --help
      • Using uvx: uvx wdoc --help
    • In any case, it is recommended to try to install pdftotext with pip11 install -U wdoc[pdftotext] as well as add fasttext support with pip11 install -U wdoc[fasttext].
  2. Add the API key for the backend you want as an environment variable: for example export OPENAI_API_KEY="***my_key***"
  3. Launch is as easy as using wdoc --task=query --path=MYDOC [ARGS] and wdoc --task=summary --path=MYDOC [ARGS] (you can use wdoc instead of wdoc)
    • If for some reason this fails, maybe try with python -m wdoc. And if everything fails, clone this repo and try again after cd inside it.
    • To get shell autocompletion: if you're using zsh: eval $(cat shell_completions/wdoc_completion.zsh). Also provided for bash and fish. You can generate your own with wdoc -- --completion MYSHELL > my_completion_file".
    • Don't forget that if you're using a lot of documents (notably via recursive filetypes) it can take a lot of time (depending on parallel processing too, but you then might run into memory errors).
  4. To ask questions about a local document: wdoc query --path="PATH/TO/YOUR/FILE" --filetype="auto"
    • If you want to reduce the startup time by directly loading the embeddings from a previous run (although the embeddings are always cached anyway): add --saveas="some/path" to the previous command to save the generated embeddings to a file and replace with --loadfrom "some/path" on every subsequent call.
  5. For more: read the documentation at wdoc --help

Scripts made with wdoc

FAQ

FAQ

  • Who is this for?

    • wdoc is for power users who want document querying on steroid, and in depth AI powered document summaries.

  • What's RAG?

    • A RAG system (retrieval augmented generation) is basically an LLM powered search through a text corpus.

  • Why make another RAG system? Can't you use any of the others?

    • I'm a medical student so I need to be able to ask medical question from a lot (tens of thousands) of documents, of different types (epub, pdf, anki database, Logseq, website dump, youtube videos and playlists, recorded conferences, audio files, etc).

  • Why is wdoc better than most RAG system to ask questions on documents?

    • It uses both a strong and query_eval LLM. After finding the appropriate documents using embeddings, the query_eval LLM is used to filter through the documents that don't seem to be about the question, then the strong LLM answers the question based on each remaining documents, then combines them all in a neat markdown. Also wdoc is very customizable.

  • Why can wdoc also produce summaries?

    • I have little free time so I needed a tailor made summary feature to keep up with the news. But most summary systems are rubbish and just try to give you the high level takeaway points, and don't handle properly text chunking. So I made my own tailor made summarizer. The summary prompts can be found in utils/prompts.py and focus on extracting the arguments/reasonning/though process/arguments of the author then use markdown indented bullet points to make it easy to read. It's really good! The prompts dataclass is not frozen so you can provide your own prompt if you want.

  • What other tasks are supported by wdoc?

  • Which LLM providers are supported by wdoc?

  • What do you use wdoc for?

    • I follow heterogeneous sources to keep up with the news: youtube, website, etc. So thanks to wdoc I can automatically create awesome markdown summaries that end up straight into my Logseq database as a bunch of TODO blocks.
    • I use it to ask technical questions to my vast heterogeneous corpus of medical knowledge.
    • I use it to query my personal documents using the --private argument.
    • I sometimes use it to summarize a documents then go straight to asking questions about it, all in the same command.
    • I use it to ask questions about entire youtube playlists.
    • Other use case are the reason I made the scripts made with wdoc section

  • What's up with the name?

    • One of my favorite character (and somewhat of a rolemodel is Winston Wolf and after much hesitation I decided WolfDoc would be too confusing and WinstonDoc sounds like something micro$oft would do. Also wd and wdoc were free, whereas doctools was already taken. The initial name of the project was DocToolsLLM, a play on words between 'doctor' and 'tool'.

  • How can I improve the prompt for a specific task without coding?

    • Each prompt of the query task are roleplaying as employees working for WDOC-CORP©, either as Eve the Evaluator (the LLM that filters out relevant documents), Anna the Answerer (the LLM that answers the question from a filtered document) or Carl the Combiner (the LLM that combines answers from Answerer as one). There's also Sam the Summarizer for summaries and Raphael the Rephraser to expand your query. They are all receiving orders from you if you talk to them in a prompt.

  • How can I use wdoc's parser for my own documents?

    • If you are in the shell cli you can easily use wdoc parse my_file.pdf (this actually replaces the call to call instead wdoc_parse_file my_file.pdf). add --only_text to only get the text and no metadata. If you're having problem with argument parsing you can try adding the --pipe argument.
    • If you want the document using python: 
      from wdoc import wdoc
list_of_docs = Wdoc.parse_file(path=my_path)

  • What should I do if my PDF are encrypted?

    • If you're on linux you can try running qpdf --decrypt input.pdf output.pdf

  • How can I add my own pdf parser?

    • Write a python class and add it there: wdoc.utils.loaders.pdf_loaders['parser_name']=parser_object then call wdoc with --pdf_parsers=parser_name.
      • The class has to take a path argument in __init__, have a load method taking no argument but returning a List[Document]. Take a look at the OpenparseDocumentParser class for an example.

  • What should I do if I keep hitting rate limits?

    • The simplest way is to add the debug argument. It will disable multithreading, multiprocessing and LLM concurrency. A less harsh alternative is to set the environment variable WDOC_LLM_MAX_CONCURRENCY to a lower value.