embedchain is a framework to easily create LLM powered bots over any dataset. If you want a javascript version, check out embedchain-js
- Latest Updates
- What is embedchain?
- Getting Started
- Advanced
- How does it work?
- Contribution Guidelines
- Tech Stack
- Team
-
Introduce a new interface called
chat
. It remembers the history (last 5 messages) and can be used to powerful stateful bots. You can use it by calling.chat
on any app instance. Works for both OpenAI and OpenSourceApp. -
Introduce a new app type called
OpenSourceApp
. It usesgpt4all
as the LLM andsentence transformers
all-MiniLM-L6-v2 as the embedding model. If you use this app, you dont have to pay for anything.
Embedchain abstracts the entire process of loading a dataset, chunking it, creating embeddings and then storing in a vector database.
You can add a single or multiple dataset using .add
and .add_local
function and then use .query
function to find an answer from the added datasets.
If you want to create a Naval Ravikant bot which has 1 youtube video, 1 book as pdf and 2 of his blog posts, as well as a question and answer pair you supply, all you need to do is add the links to the videos, pdf and blog posts and the QnA pair and embedchain will create a bot for you.
from embedchain import App
naval_chat_bot = App()
# Embed Online Resources
naval_chat_bot.add("youtube_video", "https://www.youtube.com/watch?v=3qHkcs3kG44")
naval_chat_bot.add("pdf_file", "https://navalmanack.s3.amazonaws.com/Eric-Jorgenson_The-Almanack-of-Naval-Ravikant_Final.pdf")
naval_chat_bot.add("web_page", "https://nav.al/feedback")
naval_chat_bot.add("web_page", "https://nav.al/agi")
# Embed Local Resources
naval_chat_bot.add_local("qna_pair", ("Who is Naval Ravikant?", "Naval Ravikant is an Indian-American entrepreneur and investor."))
naval_chat_bot.query("What unique capacity does Naval argue humans possess when it comes to understanding explanations or concepts?")
# answer: Naval argues that humans possess the unique capacity to understand explanations or concepts to the maximum extent possible in this physical reality.
First make sure that you have the package installed. If not, then install it using pip
pip install embedchain
Creating a chatbot involves 3 steps:
- Import the App instance (App Types)
- Add Dataset (Add Dataset)
- Query or Chat on the dataset and get answers (Interface Types)
We have three types of App.
from embedchain import App
naval_chat_bot = App()
-
App
uses OpenAI's model, so these are paid models. You will be charged for embedding model usage and LLM usage. -
App
uses OpenAI's embedding model to create embeddings for chunks and ChatGPT API as LLM to get answer given the relevant docs. Make sure that you have an OpenAI account and an API key. If you have don't have an API key, you can create one by visiting this link. -
Once you have the API key, set it in an environment variable called
OPENAI_API_KEY
import os
os.environ["OPENAI_API_KEY"] = "sk-xxxx"
from embedchain import OpenSourceApp
naval_chat_bot = OpenSourceApp()
-
OpenSourceApp
uses open source embedding and LLM model. It usesall-MiniLM-L6-v2
from Sentence Transformers library as the embedding model andgpt4all
as the LLM. -
Here there is no need to setup any api keys. You just need to install embedchain package and these will get automatically installed.
-
Once you have imported and instantiated the app, every functionality from here onwards is the same for either type of app.
from embedchain import PersonApp
naval_chat_bot = PersonApp("name_of_person_or_character") #Like "Yoda"
-
PersonApp
uses OpenAI's model, so these are paid models. You will be charged for embedding model usage and LLM usage. -
PersonApp
uses OpenAI's embedding model to create embeddings for chunks and ChatGPT API as LLM to get answer given the relevant docs. Make sure that you have an OpenAI account and an API key. If you have don't have an API key, you can create one by visiting this link. -
Once you have the API key, set it in an environment variable called
OPENAI_API_KEY
import os
os.environ["OPENAI_API_KEY"] = "sk-xxxx"
-
This step assumes that you have already created an
app
instance by either usingApp
orOpenSourceApp
. We are calling our app instance asnaval_chat_bot
-
Now use
.add
function to add any dataset.
# naval_chat_bot = App() or
# naval_chat_bot = OpenSourceApp()
# Embed Online Resources
naval_chat_bot.add("youtube_video", "https://www.youtube.com/watch?v=3qHkcs3kG44")
naval_chat_bot.add("pdf_file", "https://navalmanack.s3.amazonaws.com/Eric-Jorgenson_The-Almanack-of-Naval-Ravikant_Final.pdf")
naval_chat_bot.add("web_page", "https://nav.al/feedback")
naval_chat_bot.add("web_page", "https://nav.al/agi")
# Embed Local Resources
naval_chat_bot.add_local("qna_pair", ("Who is Naval Ravikant?", "Naval Ravikant is an Indian-American entrepreneur and investor."))
- If there is any other app instance in your script or app, you can change the import as
from embedchain import App as EmbedChainApp
from embedchain import OpenSourceApp as EmbedChainOSApp
from embedchain import PersonApp as EmbedChainPersonApp
# or
from embedchain import App as ECApp
from embedchain import OpenSourceApp as ECOSApp
from embedchain import PersonApp as ECPApp
-
This interface is like a question answering bot. It takes a question and gets the answer. It does not maintain context about the previous chats.
-
To use this, call
.query
function to get the answer for any query.
print(naval_chat_bot.query("What unique capacity does Naval argue humans possess when it comes to understanding explanations or concepts?"))
# answer: Naval argues that humans possess the unique capacity to understand explanations or concepts to the maximum extent possible in this physical reality.
-
This interface is chat interface where it remembers previous conversation. Right now it remembers 5 conversation by default.
-
To use this, call
.chat
function to get the answer for any query.
print(naval_chat_bot.chat("How to be happy in life?"))
# answer: The most important trick to being happy is to realize happiness is a skill you develop and a choice you make. You choose to be happy, and then you work at it. It's just like building muscles or succeeding at your job. It's about recognizing the abundance and gifts around you at all times.
print(naval_chat_bot.chat("who is naval ravikant?"))
# answer: Naval Ravikant is an Indian-American entrepreneur and investor.
print(naval_chat_bot.chat("what did the author say about happiness?"))
# answer: The author, Naval Ravikant, believes that happiness is a choice you make and a skill you develop. He compares the mind to the body, stating that just as the body can be molded and changed, so can the mind. He emphasizes the importance of being present in the moment and not getting caught up in regrets of the past or worries about the future. By being present and grateful for where you are, you can experience true happiness.
-
You can add config to your query method to stream responses like ChatGPT does. You would require a downstream handler to render the chunk in your desirable format. Supports both OpenAI model and OpenSourceApp.
-
To use this, instantiate a
QueryConfig
orChatConfig
object withstream=True
. Then pass it to the.chat()
or.query()
method. The following example iterates through the chunks and prints them as they appear.
app = App()
query_config = QueryConfig(stream = True)
resp = app.query("What unique capacity does Naval argue humans possess when it comes to understanding explanations or concepts?", query_config)
for chunk in resp:
print(chunk, end="", flush=True)
# answer: Naval argues that humans possess the unique capacity to understand explanations or concepts to the maximum extent possible in this physical reality.
We support the following formats:
To add any youtube video to your app, use the data_type (first argument to .add
) as youtube_video
. Eg:
app.add('youtube_video', 'a_valid_youtube_url_here')
To add any pdf file, use the data_type as pdf_file
. Eg:
app.add('pdf_file', 'a_valid_url_where_pdf_file_can_be_accessed')
Note that we do not support password protected pdfs.
To add any web page, use the data_type as web_page
. Eg:
app.add('web_page', 'a_valid_web_page_url')
To add any doc/docx file, use the data_type as docx
. Eg:
app.add('docx', 'a_local_docx_file_path')
To supply your own text, use the data_type as text
and enter a string. The text is not processed, this can be very versatile. Eg:
app.add_local('text', 'Seek wealth, not money or status. Wealth is having assets that earn while you sleep. Money is how we transfer time and wealth. Status is your place in the social hierarchy.')
Note: This is not used in the examples because in most cases you will supply a whole paragraph or file, which did not fit.
To supply your own QnA pair, use the data_type as qna_pair
and enter a tuple. Eg:
app.add_local('qna_pair', ("Question", "Answer"))
To add a XML site map containing list of all urls, use the data_type as sitemap
and enter the sitemap url. Eg:
app.add('sitemap', 'a_valid_sitemap_url/sitemap.xml')
Default behavior is to create a persistent vector DB in the directory ./db. You can split your application into two Python scripts: one to create a local vector DB and the other to reuse this local persistent vector DB. This is useful when you want to index hundreds of documents and separately implement a chat interface.
Create a local index:
from embedchain import App
naval_chat_bot = App()
naval_chat_bot.add("youtube_video", "https://www.youtube.com/watch?v=3qHkcs3kG44")
naval_chat_bot.add("pdf_file", "https://navalmanack.s3.amazonaws.com/Eric-Jorgenson_The-Almanack-of-Naval-Ravikant_Final.pdf")
You can reuse the local index with the same code, but without adding new documents:
from embedchain import App
naval_chat_bot = App()
print(naval_chat_bot.query("What unique capacity does Naval argue humans possess when it comes to understanding explanations or concepts?"))
- If you want to add any other format, please create an issue and we will add it to the list of supported formats.
Before you consume valueable tokens, you should make sure that the embedding you have done works and that it's receiving the correct document from the database.
For this you can use the dry_run
method.
Following the example above, add this to your script:
print(naval_chat_bot.dry_run('Can you tell me who Naval Ravikant is?'))
'''
Use the following pieces of context to answer the query at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer.
Q: Who is Naval Ravikant?
A: Naval Ravikant is an Indian-American entrepreneur and investor.
Query: Can you tell me who Naval Ravikant is?
Helpful Answer:
'''
The embedding is confirmed to work as expected. It returns the right document, even if the question is asked slightly different. No prompt tokens have been consumed.
The dry run will still consume tokens to embed your query, but it is only ~1/15 of the prompt.
Chinese Colab Tutorial:https://colab.research.google.com/drive/10_7Y0x4YXWVjuhhYwVraGQLpKAatTQTm?usp=sharing
Chinese Video Tutorial:https://www.bilibili.com/video/BV1YX4y1H7oN
Embedchain is made to work out of the box. However, for advanced users we're also offering configuration options. All of these configuration options are optional and have sane defaults.
Here's the readme example with configuration options.
import os
from embedchain import App
from embedchain.config import InitConfig, AddConfig, QueryConfig
from chromadb.utils import embedding_functions
# Example: use your own embedding function
config = InitConfig(ef=embedding_functions.OpenAIEmbeddingFunction(
api_key=os.getenv("OPENAI_API_KEY"),
organization_id=os.getenv("OPENAI_ORGANIZATION"),
model_name="text-embedding-ada-002"
))
naval_chat_bot = App(config)
# Example: define your own chunker config for `youtube_video`
youtube_add_config = {
"chunker": {
"chunk_size": 1000,
"chunk_overlap": 100,
"length_function": len,
}
}
naval_chat_bot.add("youtube_video", "https://www.youtube.com/watch?v=3qHkcs3kG44", AddConfig(**youtube_add_config))
add_config = AddConfig()
naval_chat_bot.add("pdf_file", "https://navalmanack.s3.amazonaws.com/Eric-Jorgenson_The-Almanack-of-Naval-Ravikant_Final.pdf", add_config)
naval_chat_bot.add("web_page", "https://nav.al/feedback", add_config)
naval_chat_bot.add("web_page", "https://nav.al/agi", add_config)
naval_chat_bot.add_local("qna_pair", ("Who is Naval Ravikant?", "Naval Ravikant is an Indian-American entrepreneur and investor."), add_config)
query_config = QueryConfig() # Currently no options
print(naval_chat_bot.query("What unique capacity does Naval argue humans possess when it comes to understanding explanations or concepts?", query_config))
Here's the example of using custom prompt template with .query
from embedchain.config import QueryConfig
from embedchain.embedchain import App
from string import Template
import wikipedia
einstein_chat_bot = App()
# Embed Wikipedia page
page = wikipedia.page("Albert Einstein")
einstein_chat_bot.add("text", page.content)
# Example: use your own custom template with `$context` and `$query`
einstein_chat_template = Template("""
You are Albert Einstein, a German-born theoretical physicist,
widely ranked among the greatest and most influential scientists of all time.
Use the following information about Albert Einstein to respond to
the human's query acting as Albert Einstein.
Context: $context
Keep the response brief. If you don't know the answer, just say that you don't know, don't try to make up an answer.
Human: $query
Albert Einstein:""")
query_config = QueryConfig(einstein_chat_template)
queries = [
"Where did you complete your studies?",
"Why did you win nobel prize?",
"Why did you divorce your first wife?",
]
for query in queries:
response = einstein_chat_bot.query(query, query_config)
print("Query: ", query)
print("Response: ", response)
# Output
# Query: Where did you complete your studies?
# Response: I completed my secondary education at the Argovian cantonal school in Aarau, Switzerland.
# Query: Why did you win nobel prize?
# Response: I won the Nobel Prize in Physics in 1921 for my services to Theoretical Physics, particularly for my discovery of the law of the photoelectric effect.
# Query: Why did you divorce your first wife?
# Response: We divorced due to living apart for five years.
Client Mode. By defining a (ChromaDB) server, you can run EmbedChain as a client only.
from embedchain import App
config = InitConfig(host="localhost", port="8080")
app = App(config)
This is useful for scalability. Say you have EmbedChain behind an API with multiple workers. If you separate clients and server, all clients can connect to the server, which only has to keep one instance of the database in memory. You also don't have to worry about replication.
To run a chroma db server, run git clone https://github.com/chroma-core/chroma.git
, navigate to the directory (cd chroma
) and then start the server with docker-compose up -d --build
.
This section describes all possible config options.
option | description | type | default |
---|---|---|---|
log_level | log level | string | WARNING |
ef | embedding function | chromadb.utils.embedding_functions | {text-embedding-ada-002} |
db | vector database (experimental) | BaseVectorDB | ChromaDB |
host | hostname for (Chroma) DB server | string | None |
port | port number for (Chroma) DB server | string, int | None |
option | description | type | default |
---|---|---|---|
chunker | chunker config | ChunkerConfig | Default values for chunker depends on the data_type . Please refer ChunkerConfig |
loader | loader config | LoaderConfig | None |
option | description | type | default |
---|---|---|---|
chunk_size | Maximum size of chunks to return | int | Default value for various data_type mentioned below |
chunk_overlap | Overlap in characters between chunks | int | Default value for various data_type mentioned below |
length_function | Function that measures the length of given chunks | typing.Callable | Default value for various data_type mentioned below |
Default values of chunker config parameters for different data_type
:
data_type | chunk_size | chunk_overlap | length_function |
---|---|---|---|
docx | 1000 | 0 | len |
text | 300 | 0 | len |
qna_pair | 300 | 0 | len |
web_page | 500 | 0 | len |
pdf_file | 1000 | 0 | len |
youtube_video | 2000 | 0 | len |
coming soon
option | description | type | default |
---|---|---|---|
number_documents | number of documents to be retrieved as context | int | 1 |
template | custom template for prompt | Template | Template("Use the following pieces of context to answer the query at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer. $context Query: $query Helpful Answer:") |
history | include conversation history from your client or database | any (recommendation: list[str]) | None |
stream | control if response is streamed back to the user | bool | False |
model | OpenAI model | string | gpt-3.5-turbo-0613 |
temperature | creativity of the model (0-1) | float | 0 |
max_tokens | limit maximum tokens used | int | 1000 |
top_p | diversity of words used by the model (0-1) | float | 1 |
All options for query and...
coming soon
history is handled automatically, the config option is not supported.
Resets the database and deletes all embeddings. Irreversible. Requires reinitialization afterwards.
app.reset()
Counts the number of embeddings (chunks) in the database.
print(app.count())
# returns: 481
Creating a chat bot over any dataset needs the following steps to happen
- load the data
- create meaningful chunks
- create embeddings for each chunk
- store the chunks in vector database
Whenever a user asks any query, following process happens to find the answer for the query
- create the embedding for query
- find similar documents for this query from vector database
- pass similar documents as context to LLM to get the final answer.
The process of loading the dataset and then querying involves multiple steps and each steps has nuances of it is own.
- How should I chunk the data? What is a meaningful chunk size?
- How should I create embeddings for each chunk? Which embedding model should I use?
- How should I store the chunks in vector database? Which vector database should I use?
- Should I store meta data along with the embeddings?
- How should I find similar documents for a query? Which ranking model should I use?
These questions may be trivial for some but for a lot of us, it needs research, experimentation and time to find out the accurate answers.
embedchain is a framework which takes care of all these nuances and provides a simple interface to create bots over any dataset.
In the first release, we are making it easier for anyone to get a chatbot over any dataset up and running in less than a minute. All you need to do is create an app instance, add the data sets using .add
function and then use .query
function to get the relevant answer.
Thank you for your interest in contributing to the EmbedChain project! We welcome your ideas and contributions to help improve the project. Please follow the instructions below to get started:
-
Fork the repository: Click on the "Fork" button at the top right corner of this repository page. This will create a copy of the repository in your own GitHub account.
-
Install the required dependencies: Ensure that you have the necessary dependencies installed in your Python environment. You can do this by running the following command:
make install
- Make changes in the code: Create a new branch in your forked repository and make your desired changes in the codebase.
- Format code: Before creating a pull request, it's important to ensure that your code follows our formatting guidelines. Run the following commands to format the code:
make lint format
- Create a pull request: When you are ready to contribute your changes, submit a pull request to the EmbedChain repository. Provide a clear and descriptive title for your pull request, along with a detailed description of the changes you have made.
embedchain is built on the following stack:
- Langchain as an LLM framework to load, chunk and index data
- OpenAI's Ada embedding model to create embeddings
- OpenAI's ChatGPT API as LLM to get answers given the context
- Chroma as the vector database to store embeddings
- gpt4all as an open source LLM
- sentence-transformers as open source embedding model
- Taranjeet Singh (@taranjeetio)
If you utilize this repository, please consider citing it with:
@misc{embedchain,
author = {Taranjeet Singh},
title = {Embechain: Framework to easily create LLM powered bots over any dataset},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/embedchain/embedchain}},
}