README_LangChain.md

h2oGPT integration with LangChain and Chroma/FAISS/Weaviate for Vector DB

Our goal is to make it easy to have private offline document question-answer using LLMs.

Getting Started

Follow the main README getting started steps. In this readme, we focus on other optional aspects.

To support GPU FAISS database, run:

pip install -r reqs_optional/requirements_optional_faiss.txt

or for CPU FAISS database, run:

pip install -r reqs_optional/requirements_optional_faiss_cpu.txt

or for Weaviate, run:

pip install -r reqs_optional/requirements_optional_langchain.txt

Supported Datatypes

Open-source data types are supported, .msg is not supported due to GPL-3 requirement. Other meta types support other types inside them. Special support for some behaviors is provided by the UI itself.

Supported Native Datatypes

• .pdf: Portable Document Format (PDF),
• .txt: Text file (UTF-8),
• .csv: CSV,
• .toml: Toml,
• .py: Python,
• .rst: reStructuredText,
• .rtf: Rich Text Format,
• .md: Markdown,
• .html: HTML File,
• .mhtml: MHTML File,
• .htm: HTML File,
• .docx: Word Document (optional),
• .doc: Word Document (optional),
• .xlsx: Excel Document (optional),
• .xls: Excel Document (optional),
• .enex: EverNote,
• .eml: Email,
• .epub: EPub,
• .odt: Open Document Text,
• .pptx : PowerPoint Document,
• .ppt : PowerPoint Document,
• .xml: XML,
• .apng : APNG Image (optional),
• .blp : BLP Image (optional),
• .bmp : BMP Image (optional),
• .bufr : BUFR Image (optional),
• .bw : BW Image (optional),
• .cur : CUR Image (optional),
• .dcx : DCX Image (optional),
• .dds : DDS Image (optional),
• .dib : DIB Image (optional),
• .emf : EMF Image (optional),
• .eps : EPS Image (optional),
• .fit : FIT Image (optional),
• .fits : FITS Image (optional),
• .flc : FLC Image (optional),
• .fli : FLI Image (optional),
• .fpx : FPX Image (optional),
• .ftc : FTC Image (optional),
• .ftu : FTU Image (optional),
• .gbr : GBR Image (optional),
• .gif : GIF Image (optional),
• .grib : GRIB Image (optional),
• .h5 : H5 Image (optional),
• .hdf : HDF Image (optional),
• .icb : ICB Image (optional),
• .icns : ICNS Image (optional),
• .ico : ICO Image (optional),
• .iim : IIM Image (optional),
• .im : IM Image (optional),
• .j2c : J2C Image (optional),
• .j2k : J2K Image (optional),
• .jfif : JFIF Image (optional),
• .jp2 : JP2 Image (optional),
• .jpc : JPC Image (optional),
• .jpe : JPE Image (optional),
• .jpeg : JPEG Image (optional),
• .jpf : JPF Image (optional),
• .jpg : JPG Image (optional),
• .jpx : JPX Image (optional),
• .mic : MIC Image (optional),
• .mpeg : MPEG Image (optional),
• .mpg : MPG Image (optional),
• .msp : MSP Image (optional),
• .pbm : PBM Image (optional),
• .pcd : PCD Image (optional),
• .pcx : PCX Image (optional),
• .pgm : PGM Image (optional),
• .png : PNG Image (optional),
• .pnm : PNM Image (optional),
• .ppm : PPM Image (optional),
• .ps : PS Image (optional),
• .psd : PSD Image (optional),
• .pxr : PXR Image (optional),
• .qoi : QOI Image (optional),
• .ras : RAS Image (optional),
• .rgb : RGB Image (optional),
• .rgba : RGBA Image (optional),
• .sgi : SGI Image (optional),
• .tga : TGA Image (optional),
• .tif : TIF Image (optional),
• .tiff : TIFF Image (optional),
• .vda : VDA Image (optional),
• .vst : VST Image (optional),
• .webp : WEBP Image (optional),
• .wmf : WMF Image (optional),
• .xbm : XBM Image (optional),
• .xpm : XPM Image (optional).

To support image captioning, on Ubuntu run:

sudo apt-get install libmagic-dev poppler-utils tesseract-ocr libtesseract-dev

and ensure in requirements_optional_langchain.txt that unstructured[local-inference] and pdf2image are installed. Otherwise, for no image support just unstructured is sufficient.

OCR is disabled by default, but can be enabled if making database via make_db.py, and then on Ubuntu run:

sudo apt-get install tesseract-ocr libtesseract-dev

and ensure you pip install pytesseract. See Tesseract documentation.

To support Microsoft Office docx, doc, xls, xlsx, on Ubuntu run:

sudo apt-get install libreoffice

In some cases unstructured by itself cannot handle URL content properly, then we will use Selenium or PlayWright as backup methods if unstructured fails. To have this be done, do:

pip install -r reqs_optional/requirements_optional_langchain.urls.txt

For Selenium, one needs to have chrome installed, e.g. on Ubuntu:

sudo bash
apt install -y unzip xvfb libxi6 libgconf-2-4
apt install -y default-jdk
curl -sS -o - https://dl-ssl.google.com/linux/linux_signing_key.pub | apt-key add
bash -c "echo 'deb [arch=amd64] http://dl.google.com/linux/chrome/deb/ stable main' >> /etc/apt/sources.list.d/google-chrome.list"
apt -y update
apt -y install google-chrome-stable  # e.g. Google Chrome 114.0.5735.198
google-chrome --version  # e.g. Google Chrome 114.0.5735.198
# visit https://chromedriver.chromium.org/downloads and download matching version
# E.g.
wget https://chromedriver.storage.googleapis.com/114.0.5735.90/chromedriver_linux64.zip
unzip chromedriver_linux64.zip
sudo mv chromedriver /usr/bin/chromedriver
sudo chown root:root /usr/bin/chromedriver
sudo chmod +x /usr/bin/chromedriver

PlayWright is disabled by default as it hangs.

Supported Meta Datatypes

• .zip : Zip File containing any native datatype,
• .urls : Text file containing new-line separated URLs (to be consumed via download).

Supported Datatypes in UI

• Files : All Native and Meta DataTypes as file(s),
• URL : Any URL (i.e. http:// or https://),
• ArXiv : Any ArXiv name (e.g. arXiv:1706.03762),
• Text : Paste Text into UI.

To support ArXiv API, do:

pip install -r reqs_optional/requirements_optional_langchain.gpllike.txt

but pymupdf is AGPL, requiring any source code be made available, which is not an issue directly for h2oGPT, but it's like GPL and too strong a constraint for general commercial use.

When pymupdf is installed, we will use PyMuPDFLoader by default to parse PDFs since it's better than PyPDFLoader and much better than PDFMinerLoader.

Adding new file types

The function file_to_doc controls the ingestion, with allowed ones listed. If one wants to add a new file type, add it to the list file_types, and then add an entry in file_to_doc() function.

Metadata is added using add_meta function, and other metadata, like chunk_id, is added after chunking. One could add a new step to add meta data to page_content to each langchain Document.

Database creation

To use some example databases (will overwrite UserData make above unless change options) and run generate after, do:

python src/make_db.py --download_some=True
python generate.py --base_model=h2oai/h2ogpt-oasst1-512-12b --load_8bit=True --langchain_mode=UserData --langchain_modes="['UserData', 'wiki', 'MyData', 'github h2oGPT', 'DriverlessAI docs']"

which downloads example databases. This obtains files from some pre-generated databases. A large Wikipedia database is also available.

To build the database first outside chatbot, then run generate after, do:

python src/make_db.py
python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b --langchain_mode=UserData

To add data to the existing database, then run generate after, do:

python src/make_db.py --add_if_exists=True
python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b --langchain_mode=UserData

By default, generate.py will load an existing UserData database and add any documents added to user_path or change any files that have changed. To avoid detecting any new files, just avoid passing --user_path=user_path, which sets it to None, i.e.:

python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b --langchain_mode=UserData

which will avoid using user_path since it is no longer passed. Otherwise when passed, any new files will be added or changed (by hash) files will be updated (delete old sources and add new sources).

If you have enough GPU memory for embedding, but not the LLM as well, then a less private mode is to use OpenAI model.

python generate.py  --inference_server=openai_chat --base_model=gpt-3.5-turbo --score_model=None --langchain_mode=LLM --langchain_modes="['LLM', 'UserData', 'MyData']"

and if you want to push image caption model to get better captions, this can be done if have enough GPU memory or if use OpenAI:

python generate.py  --inference_server=openai_chat --base_model=gpt-3.5-turbo --score_model=None --langchain_mode=LLM --langchain_modes="['LLM', 'UserData', 'MyData']" --captions_model=Salesforce/blip2-flan-t5-xl

Similar commands can be used for Azure OpenAI, e.g.

OPENAI_API_KEY=<key> python generate.py --inference_server="openai_azure_chat:<deployment_name>:<base_url>:<api_version>" --base_model=gpt-3.5-turbo --h2ocolors=False --langchain_mode=UserData

Multiple embeddings and sources

We only support one embedding at a time for each database.

So you could use src/make_db.py to make the db for different embeddings (--hf_embedding_model like gen.py, any HF model) for each collection (e.g. UserData, UserData2) for each source folders (e.g. user_path, user_path2), and then at generate.py time you can specify those different collection names in --langchain_modes and --langchain_modes and --langchain_mode_paths. For example:

python src/make_db.py --user_path=user_path --collection_name=UserData --langchain_type=shared --hf_embedding_model=hkunlp/instructor-large
python src/make_db.py --user_path=user_path2 --collection_name=UserData2 --langchain_type=shared --hf_embedding_model=sentence-transformers/all-MiniLM-L6-v2

Note that shared is default type already, but we show above to show what options are relevant if want to change them. Then run:

python generate.py --base_model='llama' --prompt_type=llama2 --score_model=None --langchain_mode='UserData' --langchain_modes=['UserData','UserData2'] --langchain_mode_paths={'UserData':'user_path','UserData2':'user_path2'} --langchain_mode_types={'UserData':'shared','UserData2':'shared'} --model_path_llama=https://huggingface.co/TheBloke/Llama-2-7b-Chat-GGUF/resolve/main/llama-2-7b-chat.Q6_K.gguf --max_seq_len=4096

or choose 13B. And watch-out for use of whitespace. For langchain_mode_paths you can pass surrounded by "'s and have spaces.

Choosing document types

import sys
sys.path.append('src')
from src.gpt_langchain import get_supported_types
non_image_types, image_types, video_types = get_supported_types()
print(non_image_types)
print(image_types)

Select types, and pass to make_db like:

python src/make_db.py --user_path="/home/jon/Downloads/demo_data" --collection_name=VAData --enable_pdf_ocr='off' --selected_file_types="['pdf', 'html', 'htm']"
python generate.py  --base_model='llama' --prompt_type=llama2 --score_model=None --langchain_mode=VAData --langchain_modes=['VAData'] --model_path_llama=https://huggingface.co/TheBloke/Llama-2-7b-Chat-GGUF/resolve/main/llama-2-7b-chat.Q6_K.gguf --max_seq_len=4096

or choose 13B.

To ensure a collection is persisted even when not using any athentication, be sure it is shared type, e.g.:

python generate.py --base_model='llama' --prompt_type=llama2 --score_model=None --max_max_new_tokens=2048 --max_new_tokens=1024 \
       --visible_tos_tab=False --visible_hosts_tab=False --visible_models_tab=False \
       --langchain_modes="['LLM','PersistData']" --langchain_mode=PersistData \
       --langchain_mode_types="{'PersistData':'shared'}" \
       --top_k_docs=-1 --max_time=360 --save_dir=save \
       --model_path_llama=https://huggingface.co/TheBloke/Llama-2-7b-Chat-GGUF/resolve/main/llama-2-7b-chat.Q6_K.gguf \
       --max_seq_len=4096

or choose 13B.

Note about Embeddings

The default embedding for GPU is instructor-large since most accurate, however it leads to excessively high scores for references due to its flat score distribution. For CPU the default embedding is all-MiniLM-L6-v2, and it has a sharp distribution of scores, so references make sense, but it is less accurate.

Note about FAISS

FAISS filtering is not supported in h2oGPT yet, ask if this is desired to be added. So subset by document does not function for FAISS.

Using Weaviate

About

Weaviate is an open-source vector database designed to scale seamlessly into billions of data objects. This implementation supports hybrid search out-of-the-box (meaning it will perform better for keyword searches).

You can run Weaviate in 5 ways:

• SaaS – with Weaviate Cloud Services (WCS).

  WCS is a fully managed service that takes care of hosting, scaling, and updating your Weaviate instance. You can try it out for free with a sandbox that lasts for 14 days.

  To set up a SaaS Weaviate instance with WCS:

  1. Navigate to Weaviate Cloud Console.
  2. Register or sign in to your WCS account.
  3. Create a new cluster with the following settings:
     • Subscription Tier – Free sandbox for a free trial, or contact [email protected] for other options.
     • Cluster name – a unique name for your cluster. The name will become part of the URL used to access this instance.
     • Enable Authentication? – Enabled by default. This will generate a static API key that you can use to authenticate.
  4. Wait for a few minutes until your cluster is ready. You will see a green tick ✔️ when it's done. Copy your cluster URL.

• Hybrid SaaS

  If you need to keep your data on-premise for security or compliance reasons, Weaviate also offers a Hybrid SaaS option: Weaviate runs within your cloud instances, but the cluster is managed remotely by Weaviate. This gives you the benefits of a managed service without sending data to an external party.

  The Weaviate Hybrid SaaS is a custom solution. If you are interested in this option, please reach out to [email protected].

• Self-hosted – with a Docker container

  To set up a Weaviate instance with Docker:

  1. Install Docker on your local machine if it is not already installed.
  2. Install the Docker Compose Plugin
  3. Download a docker-compose.yml file with this curl command:

curl -o docker-compose.yml "https://configuration.weaviate.io/v2/docker-compose/docker-compose.yml?modules=standalone&runtime=docker-compose&weaviate_version=v1.19.6"

 Alternatively, you can use Weaviate's docker compose [configuration tool](https://weaviate.io/developers/weaviate/installation/docker-compose) to generate your own `docker-compose.yml` file.

4. Run docker compose up -d to spin up a Weaviate instance.

   To shut it down, run docker compose down.

• Self-hosted – with a Kubernetes cluster

  To configure a self-hosted instance with Kubernetes, follow Weaviate's documentation.|

• Embedded - start a weaviate instance right from your application code using the client library

  This code snippet shows how to instantiate an embedded weaviate instance and upload a document:

  import weaviate
  from weaviate.embedded import EmbeddedOptions

  client = weaviate.Client(
    embedded_options=EmbeddedOptions()
  )

  data_obj = {
    "name": "Chardonnay",
    "description": "Goes with fish"
  }

  client.data_object.create(data_obj, "Wine")

Refer to the documentation for more details about this deployment method.

How To Use

Simply pass the --db_type=weaviate argument. For example:

python src/make_db.py --db_type=weaviate
python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b \
   --langchain_mode=UserData \
   --db_type=weaviate

will use an embedded weaviate instance.

If you have a weaviate instance hosted at say http://localhost:8080, then you need to define the WEAVIATE_URL environment variable before running the scripts:

WEAVIATE_URL=http://localhost:8080 python src/make_db.py --db_type=weaviate
WEAVIATE_URL=http://localhost:8080 python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b \
   --langchain_mode=UserData \
   --db_type=weaviate

Similarly, if you had set up your weaviate instance with a username and password using the OIDC Resource Owner Password flow, you will need to define the following additional environment variables:

• WEAVIATE_USERNAME: the username used for authentication
• WEAVIATE_PASSWORD: the password used for authentication
• WEAVIATE_SCOPE: optional, defaults to "offline_access"

Notes:

• Since h2oGPT is focused on privacy, connecting to weaviate via WCS is not supported as that will expose your data to a 3rd party
• Weaviate doesn't know about persistent directory throughout code, and maintains locations based upon collection name
• Weaviate doesn't support query of all metadata except via similarity search up to 10k documents, so full list of sources is not possible in h2oGPT UI for Update UI with Document(s) from DB or Show Sources from DB

Document Question-Answer FAQ

What is UserData and MyData?

UserData: Shared with anyone who is on your server. Persisted across sessions in single location for entire server. Control upload via allow_upload_to_user_data option. Useful for collaboration.

MyData: Personal space that is inaccessible if one goes into a new browser session. Useful for public demonstrations so that every instance is independent. Or useful if user is not allowed to upload to shared UserData and wants to do Q/A.

It's work in progress to add other persistent databases and to have MyData persisted across browser sessions via cookie or other authentication.

Why does the source link not work?

For links to direct to the document and download to your local machine, the original source documents must still be present on the host system where the database was created, e.g. user_path for UserData by default. If the database alone is copied somewhere else, that host won't have access to the documents. URL links like Wikipedia will still work normally on any host.

What is h2oGPT's LangChain integration like?

• PrivateGPT . By comparison, h2oGPT has:

  • UI with chats export, import, selection, regeneration, and undo
  • UI and document Q/A, upload, download, and list
  • Parallel ingest of documents, using GPUs if present for vector embeddings, with progress bar in stdout
  • Choose which specific collection
  • Choose to get response regarding all documents or specific selected document(s) out of a collection
  • Choose to chat with LLM, get one-off LLM response to a query, or talk to a collection
  • GPU support from any hugging face model for highest performance
  • Upload a many types of docs, from PDFs to images (caption or OCR), URLs, ArXiv queries, or just plain text inputs
  • Server-Client API through gradio client
  • RLHF score evaluation for every response
  • UI with side-by-side model comparisons against two models at a time with independent chat streams
  • Fine-tuning framework with QLORA 4-bit, 8-bit, 16-bit GPU fine-tuning or CPU fine-tuning

• localGPT. By comparison, h2oGPT has similar benefits as compared to localGPT. Both h2oGPT and localGPT can use GPUs for LLMs and embeddings, including latest Vicuna or WizardLM models.

• Quiver. By comparison, Quiver requires docker but also supports audio and video and currently only supports OpenAI models and embeddings.

• LM Studio. Nice control over models and llama settings, good windows installer.

• DocsGPT. More limited document support.

• GPT4-PDF-Chatbot-LangChain. Uses OpenAI, pinecone, etc. No longer maintained.

• Vault-AI but h2oGPT is fully private and open-source by not using OpenAI or pinecone.

• DB-GPT but h2oGPT is fully commercially viable by not using Vicuna (LLaMa based with GPT3.5 training data).

• ChatBox has ability to collaborate.

• Chat2DB like DB-GPT by Alibaba.

• pdfGPT like PrivateGPT but no longer maintained.

• docquery like PrivateGPT but uses LayoutLM.

• KhoJ but also access from emacs or Obsidian.

• ChatPDF but h2oGPT is open-source and private and many more data types.

• TryGloo Semantic Search and Classification.

• Cube.

• RFPBot. Confidence score, slack integration.

• Sharly but h2oGPT is open-source and private and many more data types. Sharly and h2oGPT both allow sharing work through UserData shared collection.

• ChatDoc but h2oGPT is open-source and private. ChatDoc shows nice side-by-side view with doc on one side and chat in other. Select specific doc or text in doc for question/summary.

• Casalioy with focus on air-gap with docker, otherwise like older privateGPT.

• Perplexity but h2oGPT is open-source and private, similar control over sources.

• HayStack but h2oGPT is open-source and private. Haystack is pivot to LLMs from NLP tasks, so well-developed documentation etc. But mostly LangChain clone.

• Empler but h2oGPT is open-source and private. Empler has nice AI and content control, and focuses on use cases like marketing.

• Writesonic but h2oGPT is open-source and private. Writesonic has better image/video control.

• HuggingChat Not for commercial use, uses LLaMa and GPT3.5 training data, so violates ToS.

• Bard but h2oGPT is open-source and private. Bard has better automatic link and image use.

• ChatGPT but h2oGPT is open-source and private. ChatGPT code interpreter has better image, video, etc. handling.

• ChatGPT-Next-Web like local ChatGPT.

• Bing but h2oGPT is open-source and private. Bing has excellent search queries and handling of results.

• Bearly but h2oGPT is open-source and private. Bearly focuses on creative content creation.

• Poe but h2oGPT is open-source and private. Poe also has immediate info-wall requiring phone number.

• WiseOne but h2oGPT is open-source and private. WiseOne is reading helper.

• Poet.ly or Aify but h2oGPT is open-source and private. Poet.ly focuses on writing articles.

• PDFGPT.ai but h2oGPT is open-source and private. Only PDF and on expensive side.

• BratGPT but h2oGPT is open-source and private. Focuses on uncensored chat.

• Halist but h2oGPT is open-source and private. Uses ChatGPT but does not store chats, but can already do that now with ChatGPT.

• UltimateGPT Toolkit Android plugin for ChatGPT.

• Intellibar ChatGPT on iPhone.

• GPTMana Android Plugin.

• Genie but h2oGPT is open-source and private. Focuses on legal assistant.

• ResearchAI but h2oGPT is open-source and private. Focuses on research helper with tools.

• ChatOn but h2oGPT is open-source and private. ChatOn focuses on mobile, iPhone app.

• Ask but h2oGPT is open-source and private. Similar content control.

• Petey but h2oGPT is open-source and private. Apple Watch.

• QuickGPT but h2oGPT is open-source and private. QuickGPT is ChatGPT for Whatsapp.

• Raitoai but h2oGPT is open-source and private. Raito.ai focuses on helping writers.

• AIChat but h2oGPT is open-source and private. Heavy on ads, avoid.

• AnonChatGPT but h2oGPT is open-source and private. Anonymous use of ChatGPT, i.e. no account required.

• GPTPro but h2oGPT is open-source and private. GPTPro focuses on Android.

• Rio but h2oGPT is open-source and private. Browser-based assistant.

• CommanderGPT but h2oGPT is open-source and private. CommanderGPT focuses on MAC with a few tasks like image generation, translation, youtube query, etc.

• ThreeSigma but h2oGPT is open-source and private. Focuses on research tools, nice page linking.

• LocalAI but h2oGPT has document question/answer. LocalAI has audio transcription, image generation, and a variety of models.

• LocalLLaMa but h2oGPT has UI and GPU support. LocalLLaMa is command-line focused. Like privateGPT.

• ChartGPT Focus on drawing charts.