kg-llm-interface

Langchain-powered natural language interface to RDF knowledge-graphs.

Installation

This repository uses poetry for package management. A Makefile rule is provided to install the dependencies:

make install

Configuration

Configuration variables are loaded from the .env file or environment variables. A template configuration file is provided in .env.example.

The chat configuration (config.chat.ChatConfig) uses OpenAI by default, however you can run this tool with open source LLMs using a framework such as llamafile, openllm or localGPT. When doing so, simply provide your LLM server url using openai_api_base and the model name usingmodel.

Quickstart

You can read and run the example notebook to get a quick overview of the system. The notebook supports using the Openai API and can run locally on a laptop.

To run the notebook in a containerized environment, run:

make notebook

Server

The server can be deployed as a standalone service using the script scripts/standalone_server.sh. It will start a uvicorn server on port 8001, use chromaDB in client-only mode and use an RDF file as knowledge graph. This should work for small datasets.

Pipelines

Pipelines are used to execute one-time operations for preparing data before the chat server can operate. They load their configuration from the .env file as well, but the variables can be overriden using yaml files (run with --help for more info).

Insert triples

flowchart LR
    RDF[RDF file] -->|insert_triples.py| SPARQL(SPARQL endpoint)

Loading

Insert data from an input RDF file to a SPARQL endpoint. The input file can be in any format supported by rdflib (ttl, json-ld, rdf/xlm, ...).

Location: insert_triples.py:

SPARQL configuration can be overriden by providing a yaml file following the aikg.config.sparql.SparqlConfig schema:

python insert_triples --sparql-config-path sparql.yaml

# sparql.yaml
endpoint: http://localhost:3030/ds/query
user: admin
password: admin

CLI usage: python aikg/flows/insert_triples.py

Chroma build

flowchart LR
    SPARQL(SPARQL endpoint) -->|chroma_build.py| CHROMA(ChromaDB)

Loading

Build the chromaDB index from a SPARQL endpoint.

Location: chroma_build.py:

CLI usage: python aikg/flows/chroma_build.py

Chroma and SPARQL configurations can be overriden by providing a yaml file following the aikg.config.chroma.ChromaConfig or aikg.config.sparql.SparqlConfig schemas respectively.

Containerized service

⚠️ WIP, not functional yet

The chat server can be deployed along with the front-end, SPARQL endpoint and chromaDB server using kubernetes.

sequenceDiagram
    Front-end->>+Chat server: question
    Chat server->>+ChromaDB: question
    ChromaDB -->ChromaDB: embed
    ChromaDB-->>-Chat server: ontology triples
    Chat server-->Chat server: generate query
    Chat server-->>+SPARQL endpoint: query
    SPARQL endpoint-->SPARQL endpoint: run query
    SPARQL endpoint-->>-Chat server: result
    Chat server-->>-Front-end: answer

Loading

Contributing

All contributions are welcome. New functions and classes should have associated docstrings following the numpy style guide.

The code formatting standard we use is black, with --line-length=79 to follow PEP8 recommendations. We use pytest as our testing framework. This project uses pyproject.toml to define package information, requirements and tooling configuration.

Tests can be executed with make test. Tests use testcontainers to temporarily deploy the required services.