kg-chat

Reterival over Knowledge Graphs

This folder contains two examples to demonistrate how to use langroid to build a chatbot that can answer questions about a knowledge graph. The first example is a PyPi Packages Dependency Chatbot that can answer questions about a dependency graph of a PyPi package. The second example is a CSV Chat that can answer questions about a CSV knowledge graph.

Requirements:

1. NEO4j:

This example relies on the neo4j Database. The easiest way to get access to neo4j is by creating a cloud account at Neo4j Aura. OR you can use Neo4j Docker image using this command:

docker run --rm \
    --name neo4j \
    -p 7474:7474 -p 7687:7687 \
    -e NEO4J_AUTH=neo4j/password \
    neo4j:latest

Upon creating the account successfully, neo4j will create a text file contains account settings, please provide the following information (uri, username, password, and database), while creating the constructor Neo4jChatAgentConfig. These settings can be set inside the .env file as shown in .env-template

2. Google Custom Search API Credentials needed to enable an Agent to use the GoogleSearchTool. Follow the instruction under Optional Setup Instructions to get these API credentials.

3. Visualization The package pyvis is required to enable the visualization tool VisualizeGraph. Run pip install pyvis to install this package.

1- PyPi Packages Dependency Chatbot

This example uses a DependencyGraphAgent (derived from Neo4jChatAgent). It auto-generates a neo4j knowledge-graph based on the dependency structure of a given PyPi package. You can then ask the chatbot questions about the dependency graph. This agent uses three tools in addition to those already available to Neo4jChatAgent:

• DepGraphTool to build the dependency graph for a given pkg version, using the API at DepsDev
• GoogleSearchTool to find package version and type information. It also can answer other question from the web about other aspects after obtaining the intended information from the dependency graph. For examples:
  • Is this package/version vulnerable?
  • does the dpendency use latest version for this package verion?
  • Can I upgrade this package in the dependency graph?

The Neo4jChatAgent has access to these tools/function-calls:

• GraphSchemaTool: get schema of Neo4j knowledge-graph
• CypherRetrievalTool: generate cypher queries to get information from Neo4j knowledge-graph (Cypher is the query language for Neo4j)
• VisualizeGraph: supports only visualizing the the whole dependency graph

Running the example

Run like this:

python3 examples/kg-chat/dependency_chatbot.py

DependencyAgent then will ask you to provide the name of the PyPi package. It will then the tool GoogleSearchTool to get the version of this package (you can skip this process by providing the intended version). The DependencyAgent agent will ask to confirm the version number before proceeding with constructing the dependency graph.

Finally, after constructing the dependency graph, you can ask DependencyAgent questions about the dependency graph such as these (specific package names are used here for illustration purposes, but of course you can use other names):

• what's the depth of the graph?
• what are the direct dependencies?
• any dependency on pytorch? which version?
• Is this package pytorch vunlnerable? (Note that in this case the DependencyAgent agent will consult the tool GoogleSearchTool to get an answer from the internet.)
• tell me 3 interesting things about this package or dependency graph
• what's the path between package-1 and package-2? (provide names of package-1 and -2)
• Tell me the names of all packages in the dependency graph that use pytorch.

NOTE: the dependency graph is constructed based on DepsDev API. Therefore, the Chatbot will not be able to construct the dependency graph if this API doesn't provide dependency metadata infromation.

2- CSV Chat

This example uses a CSVGraphAgent (derived from Neo4jChatAgent).

The CSVGraphAgent allows users to ask questions about a CSV file by automatically converting it into a Neo4j knowledge graph using Cypher queries. This enables capturing complex relationships that cannot be easily handled by libraries like pandas.

If the CSV knowledge graph has not been constructed beforehand, the CSVGraphAgent provides the pandas_to_kg tool/function-call to create the necessary nodes and relationships from the CSV file. Once the CSV knowledge graph is constructed, the CSVGraphAgent can answer questions related to the CSV knowledge graph. The CSVGraphAgent has access to this tool/function-call:

• PandasToKGTool: convert a pandas DataFrame into a CSV knowledge graph.

Running the example

Run like this:

python3 examples/kg-chat/csv-chat.py

The CSVGraphAgent will have a dialog with the user to determine if they need to construct the knowledge graph. If the user chooses to construct the knowledge graph, they will be prompted to provide the location of the CSV file (URL or local file).

Under the hood, the agent will:

• Attempt to clean the CSV file after parsing it as a DataFrame.
• Determine node labels and relationships.
• Create the nodes and relationships in the Neo4j knowledge graph.

After constructing the CSV knowledge graph, you can ask the CSVGraphAgent any question about the CSV knowledge graph. You can use this IMDB CSV file or you can use your own CSV file.

NOTES:

• Unlike some other CSV -> Neo4j examples out there, here we are relying on the LLM to infer nodes and relationships from the CSV file, and generate the necessary Cypher queries to create the CSV knowledge graph. This is more flexible than a hard-coded approach.
• The agent will warn you if the CSV file is too large before proceeding with constructing the CSV knowledge graph. It will also give you the option to proceed with constructing the CSV knowledge graph based on a sample of the CSV file (i.e., a specified number of rows).
• The agent uses the function _preprocess_dataframe_for_neo4j() to clean the CSV file by removing rows that have empty values. However, you can provide your own function to clean the CSV file.