Skip to content

Latest commit

 

History

History
306 lines (205 loc) · 13.4 KB

library.md

File metadata and controls

306 lines (205 loc) · 13.4 KB
Raw

Canopy Library

For most common use cases, users can simply deploy the fully-configurable Canopy service, which provides a REST API backend for your own RAG-infused Chatbot.

For advanced users, this page describes how to use canopy core library directly to implement their own custom applications.

💡 NOTE: You can also follow the quickstart Jupyter notebook

The idea behind Canopy library is to provide a framework to build AI applications on top of Pinecone as a long memory storage for you own data. Canopy library designed with the following principles in mind:

  • Easy to use: Canopy is designed to be easy to use. It is well packaged and can be installed with a single command.
  • Modularity: Canopy is built as a collection of modules that can be used together or separately. For example, you can use the chat_engine module to build a chatbot on top of your data, or you can use the knowledge_base module to directly store and search your data.
  • Extensibility: Canopy is designed to be extensible. You can easily add your own components and extend the functionality.
  • Production ready: Canopy designed to be production ready, tested, well documented, maintained and supported.
  • Open source: Canopy is open source and free to use. It built in partnership with the community and for the community.

High level architecture

class architecture

Setup

  1. set up a virtual environment (optional)
python3 -m venv canopy-env
source canopy-env/bin/activate

more about virtual environments here

  1. install the package
pip install canopy-sdk
  1. Set up the environment variables
import os

os.environ["PINECONE_API_KEY"] = "<PINECONE_API_KEY>"
os.environ["OPENAI_API_KEY"] = "<OPENAI_API_KEY>"
CLICK HERE for more information about the environment variables
Name Description How to get it?
PINECONE_API_KEY The API key for Pinecone. Used to authenticate to Pinecone services to create indexes and to insert, delete and search data Register or log into your Pinecone account in the console. You can access your API key from the "API Keys" section in the sidebar of your dashboard
OPENAI_API_KEY API key for OpenAI. Used to authenticate to OpenAI's services for embedding and chat API You can find your OpenAI API key here. You might need to login or register to OpenAI services

Quickstart

Step 1: Initialize global Tokenizer

The Tokenizer object is used for converting text into tokens, which is the basic data represntation that is used for processing.

Since manny different classes rely on a tokenizer, Canopy uses a singleton Tokenizer object which needs to be initialized once.

Before instantiating any other canopy core objects, please initialize the Tokenizer singleton:

from canopy.tokenizer import Tokenizer
Tokenizer.initialize()

Then, each time you want to use the tokenizer, you can simply instantiate a local object:

from canopy.tokenizer import Tokenizer

# no need to pass any parameters, the global tokenizer will be used
tokenizer = Tokenizer()

tokenizer.tokenize("Hello world!")
# output: ['Hello', 'world', '!']

Since the tokenizer object created here would be the same instance that you have initialized at the beginning of this subsection.

By default, the global tokenizer is initialized with OpenAITokenizer that is based on OpenAI's tiktoken library and aligned with GPT 3 and 4 models tokenization.

👉 Click here to understand how you can configure and customize the tokenizer The `Tokenizer` singleton is holding an inner `Tokenizer` object that implements `BaseTokenizer`.

You can create your own customized tokenizer by implementing a new class that derives from BaseTokenizer, then passing this class to the Tokenizer singleton during initialization. Example:

from canopy.tokenizer import Tokenizer, BaseTokenizer

class CustomTokenizer(BaseTokenizer):
    # Implement BaseToknizer's abstract methods, like `def tokenize()` etc.
    # ....
    
Tokenizer.initialize(tokenizer_class=CustomTokenizer)

When you initialize the Tokenizer singleton, you can pass init arguments to the underlying Tokenizer class. Any init argument that is expected by the underlying class's constructor, can be passed as kwarg directly to Tokenizer.initalize(). For example:

from canopy.tokenizer import Tokenizer
from canopy.tokenizer.openai import OpenAITokenizer
Tokenizer.initialize(tokenizer_class=OpenAITokenizer, model_name="gpt2")

Will initialize the global tokenizer with OpenAITokenizer and will pass the model_name parameter to the underlying tokenizer.

Step 2: Create a knowledge base

Knowledge base is an object that is responsible for storing and query your data. It holds a connection to a single Pinecone index and provides a simple API to insert, delete and search textual documents.

To create a knowledge base, you can use the following command:

from canopy.knowledge_base import KnowledgeBase

kb = KnowledgeBase(index_name="my-index")

To create a new Pinecone index and connect it to the knowledge base, you can use the create_canopy_index method:

kb.create_canopy_index()

Then, you will be able to mange the index in Pinecone console.

If you already created a Pinecone index, you can connect it to the knowledge base with the connect method:

kb.connect()

You can always verify the connection to the Pinecone index with the verify_index_connection method:

kb.verify_index_connection()

To learn more about customizing the KnowledgeBase and its inner components, see understanding knowledgebase workings section.

Step 3: Upsert and query data

To insert data into the knowledge base, you can create a list of documents and use the upsert method:

from canopy.models.data_models import Document
documents = [Document(id="1",
                      text="U2 are an Irish rock band from Dublin, formed in 1976.",
                      source="https://en.wikipedia.org/wiki/U2"),
             Document(id="2",
                      text="Arctic Monkeys are an English rock band formed in Sheffield in 2002.",
                      source="https://en.wikipedia.org/wiki/Arctic_Monkeys",
                      metadata={"my-key": "my-value"})]
kb.upsert(documents)

Now you can query the knowledge base with the query method to find the most similar documents to a given text:

from canopy.models.data_models import Query
results = kb.query([Query(text="Arctic Monkeys music genre"),
                    Query(text="U2 music genre",
                          top_k=10,
                          metadata_filter={"my-key": "my-value"})])

print(results[0].documents[0].text)
# output: Arctic Monkeys are an English rock band formed in Sheffield in 2002.

print(f"score - {results[0].documents[0].score:.4f}")
# output: score - 0.8942

Step 4: Create a context engine

Context engine is an object that responsible to retrieve the most relevant context for a given query and token budget.
The context engine first uses the knowledge base to retrieve the most relevant documents. Then, it formalizes the textual context that will be presented to the LLM. This textual context might be structured or unstructured, depending on the use case and configuration. The output of the context engine is designed to provide the LLM the most relevant context for a given query.

To create a context engine using a knowledge base, you can use the following command:

from canopy.context_engine import ContextEngine
context_engine = ContextEngine(kb)

Then, you can use the query method to retrieve the most relevant context for a given query and token budget:

import json

result = context_engine.query([Query(text="Arctic Monkeys music genre")], max_context_tokens=100)

print(json.dumps(json.loads(result.to_text()), indent=2, ensure_ascii=False))
print(f"\n# tokens in context returned: {result.num_tokens}")

output:

{
  "query": "Arctic Monkeys music genre",
  "snippets": [
    {
      "source": "https://en.wikipedia.org/wiki/Arctic_Monkeys",
      "text": "Arctic Monkeys are an English rock band formed in Sheffield in 2002."
    },
    {
      "source": "https://en.wikipedia.org/wiki/U2",
      "text": "U2 are an Irish rock band from Dublin, formed in 1976."
    }
  ]
}

# tokens in context returned: 89

By default, to handle the token budget constraint, the context engine will use the StuffingContextBuilder that will stuff as many documents as possible into the context without exceeding the token budget, by the order they have been retrieved from the knowledge base.

Go deeper TBD

Step 5: Create a chat engine

Chat engine is an object that implements end to end chat API with RAG. Given chat history, the chat engine orchestrates its underlying context engine and LLM to run the following steps:

  1. Generate search queries from the chat history
  2. Retrieve the most relevant context for each query using the context engine
  3. Prompt the LLM with the chat history and the retrieved context to generate the next response

To create a chat engine using a context, you can use the following command:

from canopy.chat_engine import ChatEngine
chat_engine = ChatEngine(context_engine)

Then, you can start chatting!

from canopy.models.data_models import UserMessage

response = chat_engine.chat(messages=[UserMessage(content="what is the genre of Arctic Monkeys band?")], stream=False)

print(response.choices[0].message.content)

# output: The genre of the Arctic Monkeys band is rock. Source: [Wikipedia](https://en.wikipedia.org/wiki/Arctic_Monkeys)

Canopy designed to be production ready and handle any conversation length and context length. Therefore, the chat engine uses internal components to handle long conversations and long contexts. By default, long chat history is truncated to the latest messages that fits the token budget. It orchestrates the context engine to retrieve context that fits the token budget and then use the LLM to generate the next response.

Go deeper TBD

Understanding KnowledgeBase workings

The knowledge base is an object that is responsible for storing and query your data. It holds a connection to a single Pinecone index and provides a simple API to insert, delete and search textual documents.

Upsert workflow

The upsert method is used to insert of update textual documents of any size into the knowledge base. For each document, the following steps are performed:

  1. The document is chunked into smaller pieces of text, each piece is called a Chunk.
  2. Each chunk is encoded into a vector representation.
  3. The vector representation is inserted into the Pinecone index along with the document text and metadata.

Query workflow

The query method is used to retrieve the most similar documents to a given query. For each query, the following steps are performed:

  1. The query is encoded into a vector representation.
  2. The vector representation is used to search the Pinecone index for the most similar vectors.
  3. The retrieved vectors are decoded into DocumentWithScore objects and returned inside a QueryResult object.

In the future, we intend to introduce a re-ranking phase to enhance the initial query outcomes.

KnowledgeBase components

The knowledge base is composed of the following components:

  • Index: A Pinecone index that holds the vector representations of the documents.
  • Chunker: A Chunker object that is used to chunk the documents into smaller pieces of text.
  • Encoder: An RecordEncoder object that is used to encode the chunks and queries into vector representations.

By default the knowledge base is initialized with OpenAIRecordEncoder which uses OpenAI embedding API to encode the text into vector representations, and MarkdownChunker which is based on a cloned version of Langchain's MarkdownTextSplitter chunker.

You can customize each component by passing any instance of Chunker or RecordEncoder to the KnowledgeBase constructor. To read more about these classes please refer to the docstrings of Chunker and RecordEncoder.