dora-back

The backend for Document Retrieval and Analysis (DoRA)

Run using Poetry and Python

How to install the dependencies

Either clone this project in VSCode or open a new codespace (if you have not been invited into another one).

The devcontainer.json should contain all the plugins needed to get going including the installation of Poetry (which may need to be done manually).

Additionally, set the FILE_PATH environment variable to where you store the PDF file and include your OpenAI API key in the OPENAI_API_KEY environment variable.

Subsequently, run poetry update in the terminal to install all the dependencies and create the environment.

Allow GPU-inference for local models

Set the CMAKE_ARGS environment variable according to the llama-cpp-python documentation

Run the Flask server for the endpoints

Make sure to set all the environment variables like:

• CHAT_MODEL_VENDOR_NAME: the name of the chat model vendor [openai, local, huggingface]
• CHAT_MODEL_NAME: the name of the chat model (e.g. gpt-turbo-3.5)
• EMBEDDING_MODEL_VENDOR_NAME: the name of the embeddings model vendor [openai, local, huggingface]
• EMBEDDING_MODEL_NAME: the name of the embeddings model (e.g. text-embedding-ada-002)
• CURRENT_ENV: the current environment [DEV, TST, PROD]. In DEV a CORS wrapper is applied to the Flask-server, but not in TST or PROD. In PROD, the server will connect to a defined remote endpoint for the Chroma Vector DB, but in DEV and TST, it will make use of a persistent client in Python.
• CHAT_MODEL_FOLDER_PATH: the path to the folder of local chat models
• EMBEDDING_MODEL_FOLDER_PATH: the path to the folder of local embedding models
• OPENAI_API_KEY: an OpenAI API key to use an OpenAI model specified in CHAT_MODEL_NAME
• CURRENT_ENV: the current environment for the Flask server; defaults to DEV
• CHUNK_SIZE: the chunk size in which to partition the chunks from the text extracted from documents; defaults to 512 tokens.
• TOP_K_DOCUMENTS: retrieve the top-k documents; defaults to the top-5 documents.
• MINIMUM_ACCURACY: the minimum accuracy for the retrieved documents (i.e. chunks of text); defaults to 0.80
• FETCH_K_DOCUMENTS: fetch k-number of documents (only applies if STRATEGY=mmr); defaults to 100
• LAMBDA_MULT: Lambda-multiplier, the lower this number (between 0 and 1) the more diverse the documents ought to be, the higher the less diverse the document selection is; defaults to 0.2
• STRATEGY: the document ranking strategy to use; for example similarity, similarity_score_threshold or mmr (default)
• LAST_N_MESSAGES: the last n messages to include from the chat history; defaults to 5.
• CHAT_MODEL_FOLDER_PATH: the folder path to store LOCAL chat models in.
• SENTENCE_TRANSFORMERS_HOME: the folder path to store LOCAL embedding models in.
• CHAT_HISTORY_CONNECTION_STRING: an SQL-connection string pointing towards a SQL-DB where chat history can be stored in. The schema will automatically be created in the database mentioned in the SQL-connection string.
• LOGGING_FILE_PATH: a file path where the logging files will be stored.
• MARIADB_USER: the user name to access the MariaDB instance with for CRUD operations
• MARIADB_ROOT_PASSWORD: the root password for the MariaDB instance
• MARIADB_PASSWORD: the password belonging to MARIADB_USER
• MARIADB_INSTANCE_URL: the URI for SQLAlchemy pointing to the MariaDB instance; it is of this format:

mariadb+mariadbconnector://${MARIADB_USER}:${MARIADB_PASSWORD}@dora-mariadb

Then run poetry run flask --app server run

Run Flask server using Docker container

Please configure the values in the Dockerfile before proceeding.

Build the Docker container using

docker build -t dora-backend --build-arg OPENAI_API_KEY=<openai_api_key> .

The --build-arg are needed to provide options for local models or API keys. Please have a look at the Dockerfile to familiarize yourself with any defaults.

Run the Docker container using:

docker run --name <container_name> -p 5000:8000 dora-back \
-e <environment_variable>=<value> \
-e <environment_variable>=<value>

You can access the server at localhost:5000. Overriding the default values for the environment variables is optional.

Removing CORS and connecting to remote Vector DB

To be able to remove the CORS wrapper and connect to a remote vector database, set the CURRENT_ENV variable to PROD.

Query the MariaDB

1. Log in to the MariaDB instance:

docker exec -it ${CONTAINER_NAME} mariadb -u ${MARIADB_USER} -D final_answer -p \
${MARIADB_PASSWORD}

2. Run the following SQL-statement for the top-5 final answers:

SELECT TOP(5) FROM final_answer;

3. To switch to the chat_history database:

\u chat_history

4. To view the top-5 chat-history items:

SELECT TOP(5) FROM chat_history;

init.sql

The purpose of this file is to set grant privileges to the user main. I have not figured out how to parameterize this.