xuede/steamship-langchain

steamship-langchain

Steamship Python Client Library For LangChain (🦜️🔗)

Steamship is the fastest way to build, ship, and use full-lifecycle language AI.

This repository contains LangChain adapters for Steamship, enabling LangChain developers to rapidly deploy their apps on Steamship to automatically get:

• Production-ready API endpoint(s)
• Horizontal scaling across dependencies / backends
• Persistent storage of app state (including caches)
• Built-in support for Authn/z
• Multi-tenancy support
• Seamless integration with other Steamship skills (ex: audio transcription)
• Usage Metrics and Logging
• And more...

Read more about Steamship and LangChain on our website.

Installing

Install via pip:

pip install steamship-langchain

Adapters

Initial support is offered for the following (with more to follow soon):

• LLMs
  • An adapter is provided for Steamship's OpenAI integration (steamship_langchain.llms.OpenAI)
  • An adapter is provided for caching LLM calls, via Steamship's Key-Value store (SteamshipCache)
• Callbacks
  • A callback that uses Python's logging module to record events is provided (steamship_langchain.callbacks.LoggingCallbackHandler). This can be used with ship logs to access verbose logs when deployed.
• Document Loaders
  • An adapter for exporting Steamship Files as LangChain Documents is provided (steamship_langchain.document_loaders.SteamshipLoader)
• Tools
  • Search:
    • An adapter is provided for Steamship's SERPAPI integration (SteamshipSERP)
• Memory
  • Chat History (steamship_langchain.memory.ChatMessageHistory)
• VectorStores
  • An adapter is provided for a persistent VectorStore (steamship_langchain.vectorstores.SteamshipVectorStore)
• Text Splitters
  • A splitter for Python code, based on the AST, is provided (steamship_langchain.python_splitter.PythonCodeSplitter). This provides additional context for code snippets (parent classes) while breaking the code into segments around function definitions.
• Miscellaneous Utilities
  • Importing data into Steamship
    • In order to take advantage of Steamship's persistent storage, an initial set of loader utilities are provided for a variety of sources, including:
      • Text files: steamship_langchain.file_loaders.TextFileLoader
      • Directories: steamship_langchain.file_loaders.DirectoryLoader
      • GitHub repositories: steamship_langchain.file_loaders.GitHubRepositoryLoader
      • Sphinx documentation sites: steamship_langchain.file_loaders.SphinxSiteLoader (and others)
      • YouTube videos: steamship_langchain.file_loaders.YouTubeFileLoader
      • Various text and image formats: steamship_langchain.file_loaders.UnstructuredFileLoader

📖 Documentation

Please see our here for full documentation on:

• Getting started (installation, setting up the environment, simple examples)
• How-To examples (demos, integrations, helper functions)

Example Use Cases

Here are a few examples of using LangChain on Steamship:

• Basic Prompting
• Self Ask With Search
• ChatBot
• Summarize Audio
• Question Answering With Sources

The examples use temporary workspaces to provide full cleanup during experimentation. Workspaces provide a unit of tenant isolation within Steamship. For production uses, persistent workspaces can be created and retrieved via Steamship(workspace_handle="my_workspace") .

NOTE These examples omit import blocks. Please consult the examples/ directory for complete source code.

NOTE Client examples assume that the user has a Steamship API key and that it is exposed to the environment (see: API Keys)

Basic Prompting

Example of a basic prompt using a Steamship LLM integration (full source: examples/greeting)

Server Snippet

from steamship_langchain.llms import OpenAI

@post("greet")
def greet(self, user: str) -> str:
    prompt = PromptTemplate(
      input_variables=["user"],
      template=
      "Create a welcome message for user {user}. Thank them for running their LangChain app on Steamship. "
      "Encourage them to deploy their app via `ship it` when ready.",
    )
    llm = OpenAI(client=self.client, temperature=0.8)
    return llm(prompt.format(user=user))

Client Snippet

with Steamship.temporary_workspace() as client:
    api = client.use("my-langchain-app")
    while True:
        name = input("Name: ")
        print(f'{api.invoke("/greet", user=name).strip()}\n')

Self Ask With Search

Executes the LangChain self-ask-with-search agent using the Steamship GPT and SERP Tool plugins (full source: examples/self-ask-with-search)

Server Snippet

from steamship_langchain.llms import OpenAI

@post("/self_ask_with_search")
def self_ask_with_search(self, query: str) -> str:
    llm = OpenAI(client=self.client, temperature=0.0, cache=True)
    serp_tool = SteamshipSERP(client=self.client, cache=True)
    tools = [Tool(name="Intermediate Answer", func=serp_tool.search)]
    self_ask_with_search = initialize_agent(tools, llm, agent="self-ask-with-search", verbose=False)
    return self_ask_with_search.run(query)

Client Snippet

with Steamship.temporary_workspace() as client:
    api = client.use("my-langchain-app")
    query = "Who was president the last time the Twins won the World Series?"
    print(f"Query: {query}")
    print(f"Answer: {api.invoke('/self_ask_with_search', query=query)}")

ChatBot

Implements a basic Chatbot (similar to ChatGPT) in Steamship with LangChain (full source: examples/chatbot).

NOTE The full ChatBot transcript will persist for the lifetime of the Steamship Workspace.

Server Snippet

from langchain.memory import ConversationBufferWindowMemory

from steamship_langchain.llms import OpenAIChat
from steamship_langchain.memory import ChatMessageHistory

@post("/send_message")
def send_message(self, message: str, chat_history_handle: str) -> str:
  chat_memory = ChatMessageHistory(client=self.client, key=chat_history_handle)
  mem = ConversationBufferWindowMemory(chat_memory=chat_memory, k=2)
  chatgpt = LLMChain(
    llm=OpenAIChat(client=self.client, temperature=0),
    prompt=CHATBOT_PROMPT,
    memory=mem,
  )

  return chatgpt.predict(human_input=message)

Client Snippet

with Steamship.temporary_workspace() as client:
    api = client.use("my-langchain-app")
    session_handle = "foo-user-session-1234"
    while True:
        msg = input("You: ")
        print(f"AI: {api.invoke('/send_message', message=msg, chat_history_handle=session_handle)}")

Summarize Audio (Async Chaining)

Audio transcription support not yet considered fully-production ready on Steamship. We are working hard on productionizing support for audio transcription at scale, but there may be some existing issues that you encounter as you try this out.

This provides an example of using LangChain to process audio transcriptions obtained via Steamship's speech-to-text plugins (full source: examples/summarize-audio)

A brief introduction to the Task system (and Task dependencies, for chaining) is provided in this example. Here, we use task.wait() style polling, but time-based task.refresh() style polling, etc., is also available.

Server Snippet

from steamship_langchain.llms import OpenAI

@post("summarize_file")
def summarize_file(self, file_handle: str) -> str:
    file = File.get(self.client, handle=file_handle)
    text_splitter = CharacterTextSplitter()
    texts = []
    for block in file.blocks:
        texts.extend(text_splitter.split_text(block.text))
    docs = [Document(page_content=t) for t in texts]
    llm = OpenAI(client=self.client, cache=True)
    chain = load_summarize_chain(llm, chain_type="map_reduce")
    return chain.run(docs)

@post("summarize_audio_file")
def summarize_audio_file(self, audio_file_handle: str) -> Task[str]:
    transcriber = self.client.use_plugin("whisper-s2t-blockifier")
    audio_file = File.get(self.client, handle=audio_file_handle)
    transcribe_task = audio_file.blockify(plugin_instance=transcriber.handle)
    return self.invoke_later("summarize_file", wait_on_tasks=[transcribe_task], arguments={"file_handle": audio_file.handle})

Client Snippet

churchill_yt_url = "https://www.youtube.com/watch?v=MkTw3_PmKtc"

with Steamship.temporary_workspace() as client:
    api = client.use("my-langchain-app")
    yt_importer = client.use_plugin("youtube-file-importer")
    import_task = File.create_with_plugin(client=client,
                                         plugin_instance=yt_importer.handle, 
                                         url=churchill_yt_url)
    import_task.wait()
    audio_file = import_task.output
    
    summarize_task_response = api.invoke("/summarize_audio_file", audio_file_handle=audio_file.handle)
    summarize_task = Task(client=client, **summarize_task_response)
    summarize_task.wait()
    
    if summarize_task.state == TaskState.succeeded:
      summary = base64.b64decode(summarize_task.output).decode("utf-8")
      print(f"Summary: {summary.strip()}")

Question Answering with Sources (Embeddings)

Provides a basic example of using Steamship to manage embeddings and power a LangChain agent for question answering with sources (full source: examples/qa_with_sources)

NOTE The embeddings will persist for the lifetime of the Workspace.

Server Snippet

from steamship_langchain.llms import OpenAI

def __init__(self, **kwargs):
    super().__init__(**kwargs)
    langchain.llm_cache = SteamshipCache(self.client)
    self.llm = OpenAI(client=self.client, temperature=0, cache=True, max_words=250)
    # create a persistent embedding store  
    self.index = SteamshipVectorStore(
        client=self.client, index_name="qa-demo", embedding="text-embedding-ada-002"
    )

@post("index_file")
def index_file(self, file_handle: str) -> bool:
    text_splitter = CharacterTextSplitter(chunk_size=250, chunk_overlap=0)
    file = File.get(self.client, handle=file_handle)
    texts = [text for block in file.blocks for text in text_splitter.split_text(block.text)]
    metadatas = [{"source": f"{file.handle}-offset-{i * 250}"} for i, text in enumerate(texts)]

    self.index.add_texts(texts=texts, metadatas=metadatas)
    return True

@post("search_embeddings")
def search_embeddings(self, query: str, k: int) -> List[SearchResult]:
    """Return the `k` closest items in the embedding index."""
    search_results = self.index.search(query, k=k)
    search_results.wait()
    items = search_results.output.items
    return items


@post("/qa_with_sources")
def qa_with_sources(self, query: str) -> Dict[str, Any]:
    chain = VectorDBQAWithSourcesChain.from_chain_type(
        OpenAI(client=self.client, temperature=0),
        chain_type="map_reduce",
        vectorstore=self.index,
    )

    return chain({"question": query}, return_only_outputs=False)

Client Snippet

with Steamship.temporary_workspace() as client:
    api = client.use("my-langchain-app")
    
    # Upload the State of the Union address
    with open("state_of_the_union.txt") as f:
        sotu_file = File.create(client, blocks=[Block(text=f.read())])

    # Embed
    api.invoke("/index_file", file_handle=sotu_file.handle)

    # Issue Query
    query = "What did the president say about Justice Breyer?"
    response = api.invoke("/qa_with_sources", query=query)
    print(f"Answer: {response['result'].strip()}")

API Keys

Steamship API Keys provide access to our SDK for AI models, including OpenAI, GPT, Cohere, Whisper, and more.

Get your free API key here: https://steamship.com/account/api.

Once you have an API Key, you can :

• Set the env var STEAMSHIP_API_KEY for your client
• Pass it directly via Steamship(api_key=) or Steamship.tempory_workspace(api_key=).

Alternatively, you can run ship login, which will guide you through setting up your environment.

Deploying on Steamship

Deploying LangChain apps on Steamship is simple: ship it.

From your package directory (where your api.py lives), you can issue the ship it command to generate a manifest file and push your package to Steamship. You may then use the Steamship SDK to create instances of your package in Workspaces as best fits your needs.

More on deployment and Workspaces can be found in our docs.

Feedback and Support

Have any feedback on this package? Or on Steamship in general?

We'd love to hear from you. Please reach out to: hello@steamship.com, or join us on our Discord.