Notebook

Building Semantic Memory with Embeddings¶

So far, we've mostly been treating the kernel as a stateless orchestration engine. We send text into a model API and receive text out.

In a previous notebook, we used context variables to pass in additional text into prompts to enrich them with more context. This allowed us to create a basic chat experience.

However, if you solely relied on context variables, you would quickly realize that eventually your prompt would grow so large that you would run into a the model's token limit. What we need is a way to persist state and build both short-term and long-term memory to empower even more intelligent applications.

To do this, we dive into the key concept of Semantic Memory in the Semantic Kernel.

In [ ]:

!python -m pip install semantic-kernel==0.3.10.dev0

In [ ]:

from typing import Tuple

import semantic_kernel as sk
from semantic_kernel.connectors.ai.open_ai import OpenAIChatCompletion, OpenAITextEmbedding, AzureChatCompletion, AzureTextEmbedding

In order to use memory, we need to instantiate the Kernel with a Memory Storage and an Embedding service. In this example, we make use of the VolatileMemoryStore "which can be thought of as a temporary in-memory storage (not to be confused with Semantic Memory). This memory is not written to disk and is only available during the app session.

When developing your app you will have the option to plug in persistent storage like Azure Cosmos Db, PostgreSQL, SQLite, etc. Semantic Memory allows also to index external data sources, without duplicating all the information, more on that later.

In [ ]:

kernel = sk.Kernel()

useAzureOpenAI = False

# Configure AI service used by the kernel
if useAzureOpenAI:
    deployment, api_key, endpoint = sk.azure_openai_settings_from_dot_env()
    kernel.add_chat_service("chat_completion", AzureChatCompletion(deployment, endpoint, api_key))
    # next line assumes embeddings deployment name is "text-embedding-ada-002", adjust this if  appropriate 
    kernel.add_text_embedding_generation_service("ada", AzureTextEmbedding("text-embedding-ada-002", endpoint, api_key))
else:
    api_key, org_id = sk.openai_settings_from_dot_env()
    kernel.add_chat_service("chat-gpt", OpenAIChatCompletion("gpt-3.5-turbo", api_key, org_id))
    kernel.add_text_embedding_generation_service("ada", OpenAITextEmbedding("text-embedding-ada-002", api_key, org_id))

kernel.register_memory_store(memory_store=sk.memory.VolatileMemoryStore())
kernel.import_skill(sk.core_skills.TextMemorySkill())

At its core, Semantic Memory is a set of data structures that allow you to store the meaning of text that come from different data sources, and optionally to store the source text too. These texts can be from the web, e-mail providers, chats, a database, or from your local directory, and are hooked up to the Semantic Kernel through data source connectors.

The texts are embedded or compressed into a vector of floats representing mathematically the texts' contents and meaning. You can read more about embeddings here.

Manually adding memories¶

Let's create some initial memories "About Me". We can add memories to our VolatileMemoryStore by using SaveInformationAsync

In [ ]:

async def populate_memory(kernel: sk.Kernel) -> None:
    # Add some documents to the semantic memory
    await kernel.memory.save_information_async(
        "aboutMe", id="info1", text="My name is Andrea"
    )
    await kernel.memory.save_information_async(
        "aboutMe", id="info2", text="I currently work as a tour guide"
    )
    await kernel.memory.save_information_async(
        "aboutMe", id="info3", text="I've been living in Seattle since 2005"
    )
    await kernel.memory.save_information_async(
        "aboutMe", id="info4", text="I visited France and Italy five times since 2015"
    )
    await kernel.memory.save_information_async(
        "aboutMe", id="info5", text="My family is from New York"
    )

Let's try searching the memory:

In [ ]:

async def search_memory_examples(kernel: sk.Kernel) -> None:
    questions = [
        "what's my name",
        "where do I live?",
        "where's my family from?",
        "where have I traveled?",
        "what do I do for work",
    ]

    for question in questions:
        print(f"Question: {question}")
        result = await kernel.memory.search_async("aboutMe", question)
        print(f"Answer: {result[0].text}\n")

Let's now revisit the our chat sample from the previous notebook. If you remember, we used context variables to fill the prompt with a history that continuously got populated as we chatted with the bot. Let's add also memory to it!

This is done by using the TextMemorySkill which exposes the recall native function.

recall takes an input ask and performs a similarity search on the contents that have been embedded in the Memory Store and returns the most relevant memory.

In [ ]:

async def setup_chat_with_memory(
    kernel: sk.Kernel,
) -> Tuple[sk.SKFunctionBase, sk.SKContext]:
    sk_prompt = """
    ChatBot can have a conversation with you about any topic.
    It can give explicit instructions or say 'I don't know' if
    it does not have an answer.

    Information about me, from previous conversations:
    - {{$fact1}} {{recall $fact1}}
    - {{$fact2}} {{recall $fact2}}
    - {{$fact3}} {{recall $fact3}}
    - {{$fact4}} {{recall $fact4}}
    - {{$fact5}} {{recall $fact5}}

    Chat:
    {{$chat_history}}
    User: {{$user_input}}
    ChatBot: """.strip()

    chat_func = kernel.create_semantic_function(sk_prompt, max_tokens=200, temperature=0.8)

    context = kernel.create_new_context()
    context["fact1"] = "what is my name?"
    context["fact2"] = "where do I live?"
    context["fact3"] = "where's my family from?"
    context["fact4"] = "where have I traveled?"
    context["fact5"] = "what do I do for work?"

    context[sk.core_skills.TextMemorySkill.COLLECTION_PARAM] = "aboutMe"
    context[sk.core_skills.TextMemorySkill.RELEVANCE_PARAM] = 0.8

    context["chat_history"] = ""

    return chat_func, context

The RelevanceParam is used in memory search and is a measure of the relevance score from 0.0 to 1.0, where 1.0 means a perfect match. We encourage users to experiment with different values.

Now that we've included our memories, let's chat!

In [ ]:

async def chat(
    kernel: sk.Kernel, chat_func: sk.SKFunctionBase, context: sk.SKContext
) -> bool:
    try:
        user_input = input("User:> ")
        context["user_input"] = user_input
        print(f"User:> {user_input}")
    except KeyboardInterrupt:
        print("\n\nExiting chat...")
        return False
    except EOFError:
        print("\n\nExiting chat...")
        return False

    if user_input == "exit":
        print("\n\nExiting chat...")
        return False

    answer = await kernel.run_async(chat_func, input_vars=context.variables)
    context["chat_history"] += f"\nUser:> {user_input}\nChatBot:> {answer}\n"

    print(f"ChatBot:> {answer}")
    return True

In [ ]:

print("Populating memory...")
await populate_memory(kernel)

print("Asking questions... (manually)")
await search_memory_examples(kernel)

print("Setting up a chat (with memory!)")
chat_func, context = await setup_chat_with_memory(kernel)

print("Begin chatting (type 'exit' to exit):\n")
chatting = True
while chatting:
    chatting = await chat(kernel, chat_func, context)

Adding documents to your memory¶

Many times in your applications you'll want to bring in external documents into your memory. Let's see how we can do this using our VolatileMemoryStore.

Let's first get some data using some of the links in the Semantic Kernel repo.

In [ ]:

github_files ={}
github_files["https://github.com/microsoft/semantic-kernel/blob/main/README.md"] = \
    "README: Installation, getting started, and how to contribute"
github_files["https://github.com/microsoft/semantic-kernel/blob/main/dotnet/notebooks/02-running-prompts-from-file.ipynb"] = \
    "Jupyter notebook describing how to pass prompts from a file to a semantic skill or function"
github_files["https://github.com/microsoft/semantic-kernel/blob/main/dotnet/notebooks/00-getting-started.ipynb"] = \
    "Jupyter notebook describing how to get started with the Semantic Kernel"
github_files["https://github.com/microsoft/semantic-kernel/tree/main/samples/skills/ChatSkill/ChatGPT"] = \
    "Sample demonstrating how to create a chat skill interfacing with ChatGPT"
github_files["https://github.com/microsoft/semantic-kernel/blob/main/dotnet/src/SemanticKernel/Memory/Volatile/VolatileMemoryStore.cs"] = \
    "C# class that defines a volatile embedding store"
github_files["https://github.com/microsoft/semantic-kernel/tree/main/samples/dotnet/KernelHttpServer/README.md"] = \
    "README: How to set up a Semantic Kernel Service API using Azure Function Runtime v4"
github_files["https://github.com/microsoft/semantic-kernel/tree/main/samples/apps/chat-summary-webapp-react/README.md"] = \
    "README: README associated with a sample starter react-based chat summary webapp"

Now let's add these files to our VolatileMemoryStore using SaveReferenceAsync. We'll separate these memories from the chat memories by putting them in a different collection.

In [ ]:

memory_collection_name = "SKGitHub"
print("Adding some GitHub file URLs and their descriptions to a volatile Semantic Memory.");
i = 0
for entry, value in github_files.items():
    await kernel.memory.save_reference_async(
        collection=memory_collection_name,
        description=value,
        text=value,
        external_id=entry,
        external_source_name="GitHub"
    )
    i += 1
    print("  URL {} saved".format(i))

In [ ]:

ask = "I love Jupyter notebooks, how should I get started?"
print("===========================\n" + "Query: " + ask + "\n")

memories = await kernel.memory.search_async(memory_collection_name, ask, limit=5, min_relevance_score=0.77)

i = 0
for memory in memories:
    i += 1
    print(f"Result {i}:")
    print("  URL:     : " + memory.id)
    print("  Title    : " + memory.description)
    print("  Relevance: " + str(memory.relevance))
    print()

Now you might be wondering what happens if you have so much data that it doesn't fit into your RAM? That's where you want to make use of an external Vector Database made specifically for storing and retrieving embeddings.

Stay tuned for that!