An Unofficial Python SDK and CLI for Vectara's RAG platform

Installation

pip install vektara # Stable
# OR
pip install "git+https://github.com/forrestbao/vektara.git" # Nightly
# OR
# Development mode where local changes are reflected immediately
git clone https://github.com/forrestbao/vektara.git
pip install -e .

Hello, world!

Here are the basic steps in RAG:

1. Creating a corpus, a collection of documents.
2. Ingestion: upload documents to a corpus.
3. Querying: ask questions to the corpus.

from vektara import Vectara

client = Vectara() # get credentials from environment variables

corpus_id = client.create_corpus('founding documents of the US')

client.upload(corpus_id, './test_data/US_Constition.txt')
client.upload(corpus_id, './test_data/Declaration_of_Independence.txt')

client.query(corpus_id, 'What should I do if the government becomes unjust?')

Before you start: credentials

Obtaining Vectara credentials

This client supports authentication in both Vectara's Personal API key and OAuth2. In either case, you need to obtain customer ID as well.

Setting up credentials

Following the convention of OpenAI's client, we recommend setting credentials as environment variables. You can save them as a bash script and simply source it before using the SDK.

export VECTARA_CUSTOMER_ID=123 # Required regardless of the authentication method

# For using Personal API Key to authenticate
export VECTARA_API_KEY=abc

# For using OAuth2 to authenticate
export VECTARA_CLIENT_ID=abc
export VECTARA_CLIENT_SECRET=xyz

# If both API key and OAuth2 credentials are set, the API key supercedes OAuth2.

# optional, if you are using a proxy like LlaMasterKey https://github.com/TexteaInc/LlaMasterKey/
export VECTARA_BASE_URL="http://127.0.0.1:8000/vectara"
export VECTARA_PROXY_MODE=true # Enable proxy mode

Alternatively, you can pass in your credentials as arguments when initializing the client.

Using the Python SDK

Here are the resources to learn the Python SDK:

1. Reference manual
2. Crash course
3. Demos

The SDK supports the following operations. Detailed usages can be found in the reference manual.

1. Create a corpus
2. Reset a corpus (cleans out documents in a corpus but keeps the corpus and metadata)
3. List the documents in a corpus
4. Add documents to a corpus
   • From local file(s)/folder
   • From a list of chunks (str) without any hierarchy
   • From a nested list of sections hierarchically but you don't control chunking - a section is a list of str's or sections.
5. Query a corpus
6. Set filters for a corpus which is a job in a queue
7. List a job (e.g., filter setting) in a corpus

Using the CLI client

The features in the CLI client are similar to the SDK. To learn the command line usage, run vectara --help.

You must obtain and set up your Vectara credentials as environment variables before using the command line interface.

# create a corpus
vektara create_corpus 'my knowledge base'
# output: corpus_id = 12

# upload a file to the corpus
vektara upload 12 one_file.pdf # corpurs_id = 12

# upload a folder to the corpus
vektara upload 12 ./a_folder_of_documents # corpurs_id = 12

# query the corpus
vektara query 12 'Vectara allows me to search for anything, right?' --top_k=5  # corpurs_id = 12

# reset the corpus
vektara reset_corpus 12 # corpurs_id = 12

The GUI client via Funix

The SDK can be converted into a web interface via Funix. You can drag and drop to add a file to your Vectara corpus.

pip install funix
funix src/vektara/__init__.py

Then you can access the web interface at http://localhost:3000 (the port number maybe different if port 3000 is occupied).

Below please find the screenshots.

Known bugs: Funix seems to have some memory leakage issues that the web interface may freeze after uploading a file. If that happens, please kill the process and restart the web interface.

Stylish query results

The query results are typeset Markdown ready to be rendered. A query result includes the following info:

• A summary with citations and matching scores with respect to the query
• References cited by the summary

### Here is the answer
To rearrange objects, you can utilize the "direction" attribute in a Funix decorator [1]. Manually resizing and positioning objects can be a tedious and inefficient process [2]. Another approach is to use a collision-free algorithm for auto-layout, where scopes will be resized to fit the objects inside [4]. An example of arranging objects in a column-reverse direction can be seen in the ChatGPT multiturn app [3]. Additionally, organizing your canvas with scopes can help in rearranging objects effectively [5]. Remember to experiment with these methods to find the best arrangement for your specific needs.

### References:

1. From document **funix.md** (matchness=0.673933):
  _...You can change their order and orientation using the "direction" attribute in a Funix decorator...._

2. From document **codepod.md** (matchness=0.65305215):
  _...It is painful and inefficient to resize and position the pods and scopes manually...._

3. From document **funix.md** (matchness=0.6520513):
  _...The example below shall be self-explaining:
A more advanced example is our ChatGPT multiturn app where "direction = "column-reverse"" so the message you type stays at the bottom...._

4. From document **codepod.md** (matchness=0.6495899):
  _...default}alt="Example banner"width="600"/>
After auto-layout, the pods and scopes are organized by a collision-free algorithm, and the scopes will be resized to fit the pods inside...._

5. From document **codepod.md** (matchness=0.6460015):
  _...Organize your Canvas with scopes..._

This SDK/CLI/GUI vs. Vectara's official RESTFul API

• Type less and more done. No boilerplate code.
• Copy-and-pastable examples and Jupyter notebooks to jumpstart you.
• Forget about low-level details, e.g., all metadata fields are automatically set to filterable -- under construction.
• More ways to interact
  • Command line interface (CLI)
  • GUI powered by [Funix.io]http://Funix.io) for quickly building web apps that ordinary people can use.
• More features:
  • Upload an entire folder.
  • Stylish Markdown printout for query response (see demo_simple.ipynb).
  • Log user feedback from the GUI in a local SQLite database for evaluating the quality of search and RAG.
  • Pairable with LlamaKey.ai or any API router to manage API keys and throttle requests.

Questions

Contact forrest at vectara dot com

Disclaimer

This is an UNofficial SDK and CLI for Vectara's RAG platform. Use at your own risk. Vectara does NOT provide support for this SDK or CLI.