Skip to content

Latest commit

 

History

History
143 lines (104 loc) · 8.18 KB

README.md

File metadata and controls

143 lines (104 loc) · 8.18 KB
Raw

Samples

This repository is sample graph data for developer. Samples for end user is here.

Interview

Source code: Interview

This sample graph was designed to simulate an interview with a famous person. The flow involves multiple agents working together to process user input, generate conversational context, and manage an iterative chat process.

  1. The user is prompted to input the name of a famous person through the name node.
  2. The context node generates the interview context using the input name.
  3. The messages node initializes the conversation with system and greeting messages.
  4. The chat node manages the conversation flow using a nested graph, iterating 6 times.
  5. Within the nested graph:
    • The groq node generates a response from the model.
    • The output node formats the response and prints it to the console.
    • The reducer node appends the new message to the messages array.
    • The swappedContext node swaps the roles of interviewer and interviewee.
    • The swappedMessages node updates the messages for the next iteration.
  6. The cycle repeats for 6 iterations, simulating a back-and-forth interview.

Interview in Japanese

Source code: Interview JP

  1. The user inputs the name of the person to interview.
  2. The context for the interview is generated.
  3. Initial messages are set up.
  4. The nested graph executes 6 iterations, simulating a back-and-forth interview:
    • The model generates a response.
    • The response is translated to Japanese.
    • The response is formatted and output.
    • The messages are updated with the new response.
    • The roles of interviewer and interviewee are swapped.
    • The context and messages are updated.
  5. After the iterations, the final response is translated to Japanese and output.

Chat

Source code: Chat

This sample is an interactive chat loop, where the user can engage in a conversation with an AI model. The conversation continues in a loop until the user inputs "/bye".

  1. The loop starts with continue set to true.
  2. The userInput node prompts the user for input.
  3. The checkInput node checks if the user input is "/bye". If it is, continue is set to false, breaking the loop.
  4. The userMessage node formats the user input as a message.
  5. The appendedMessages node appends the user message to the conversation history.
  6. The groq node sends the updated conversation history to the AI model for a response.
  7. The output node formats and displays the AI's response.
  8. The reducer node appends the AI's response to the conversation history.
  9. The loop repeats until the user inputs "/bye".

In-memory RAG

Souce Code: Wikipedia

This sample is a Retrieval-Augmented Generation (RAG) application using GraphAI. This graph is designed to query information about Sam Bankman-Fried's final court sentence, retrieve relevant data from Wikipedia, process it, and finally generate a response using a Large Language Model (LLM) such as GPT-3.5.

  1. The source node provides the initial data.
  2. The wikipedia node retrieves content from Wikipedia using the name from the source node.
  3. The content is split into chunks (chunks), and embeddings are generated for these chunks (embeddings). Simultaneously, an embedding for the topic is generated (topicEmbedding).
  4. The similarityCheck node calculates the similarity between the text chunks and the topic embedding.
  5. The text chunks are sorted based on similarity (sortedChunks), and the most relevant chunks are concatenated up to a token limit (referenceText).
  6. A prompt is generated using the initial query and the reference text (prompt).
  7. The RagQuery node sends the prompt to the LLM. Additionally, a one-shot query is performed using the initial query (OneShotQuery).

Research

Source code: Research

In this sample, a user inputs a topic, the system detects the language of the input, retrieves relevant information from Wikipedia, summarizes it, and then translates it if necessary.

  1. The topic node prompts the user to input a topic.
  2. The detector node identifies the language and translates the topic to English if necessary.
  3. The wikipedia node fetches and summarizes relevant information from Wikipedia.
  4. The translate node ensures the summary is in the original language if it was not in English.
  5. The result from the translate node is the final output of the workflow.

Reception

Source code: Reception

The goal of this application is to interactively gather specific information (name, date of birth, and gender) from a user through iterative conversations with a language model (LLM).

  1. The system message provides instructions to the LLM on the information to collect.
  2. The userInput node prompts the user for input.
  3. The userMessage node formats the user input.
  4. The appendedMessages node updates the conversation with the user’s message.
  5. The llm node sends the conversation to the LLM for processing.
  6. The argumentsParser node extracts information if the LLM calls the report function.
  7. The output node displays the LLM’s response.
  8. The reducer node updates the conversation with the LLM’s response.
  9. The continue node checks if the loop should continue based on the LLM’s response.

Generated Graph Example

Source code: Meta Chat

This sample aims to generate another GraphAI graph that collects user information (Name, Date of Birth, and Gender) using an LLM (GPT-4) and then executes the generated graph.

  1. The graphGenerator node is essentially prompting the GPT-4 model to generate a GraphAI graph based on the provided documentation and instructions.
  2. The parser node parses the JSON content generated by the GPT-4 model, which is the new GraphAI graph.
  3. The executer node executes the newly generated GraphAI graph by nesting it within the current graph.

Graph Description

Source code: Describe Graph

  1. The describer node asks the GPT-4 to generate the description of the specified graph.
  2. The description node extract the description from the generated message.

RSS Reader

Source code: RSS Reader

This sample fetches and processes RSS feed data from a given URL, specifically from "https://www.theverge.com/microsoft/rss/index.xml". The flow extracts relevant content from the RSS feed, processes it through a series of agents, and translates the content into Japanese.

  1. The url node provides the RSS feed URL.
  2. The rssFeed node fetches the RSS feed data from the URL.
  3. The entries node filters the fetched data to include only the required properties (title, link, content).
  4. The map node processes each filtered entry concurrently (up to 4 entries), executing the nested graph for each entry:
    • The template node formats the entry's title and content.
    • The query node translates the formatted string into Japanese.
    • The extractor node outputs the translated content.

Weather app

Source code: Weather

THis sample simulates a conversation with a meteorologist. The system is capable of handling user inputs, querying a weather API, and managing the conversation iteratively based on user requests.

  1. The loop starts with the continue node set to true.
  2. The user is prompted to input a location.
  3. The checkInput node determines if the user wants to continue or exit.
  4. If the user input is valid, it is added to the conversation messages.
  5. The llmCall node processes the messages and calls the LLM.
  6. The output node prints the LLM's response if available.
  7. If the LLM requests a tool call, the nested graph tool_calls handles the request.
  8. The reducer node updates the conversation messages.
  9. The loop continues until the user inputs "/bye".