Instructions to Run

Prerequisites

1. Python: Ensure you have Python 3.10 or higher installed on your machine. You can download it from python.org.

2. Pip: Ensure you have pip installed. It usually comes with Python. You can verify by running pip --version in your terminal.

Setup

1. Clone the Repository: Download the project files from the repository or clone it using:

   git clone https://github.com/shandilyabh/silo.small
   cd silo.small

2. Install the dependencies:

   pip install -r requirements.txt

3. Get Gemini API:

• go to: https://console.cloud.google.com/home/dashboard and sign in
• click on select project: adjacent to the google cloud branding beside the search bar
• click on new project: on the left of the pop up screen
• enter details and press Create project
• go to: https://aistudio.google.com/app/apikey
• click on create API Key
• select project
• copy API Key

4. Set up .env file: In the terminal with silo.small as current directory:

   touch .env

   open .env file and paste the gemini api key under the variable name: GEMINI_API_KEY

5. Run Silo:

   python main.py /path/to/pdf

   you can add a file's path by just dragging it into the terminal

   test run: Since, Corpus.pdf comes with the repository

   python main.py Corpus.pdf

Notes

• on the first run, Sentence Transformers will fetch Supabase's gte-small model for generating word embeddings.

Architectural Decisions:

silo uses:

• Unstructured Library for parsing the PDF and extracting text from it.
• Langchain's RecursiveCharachterTextSplitter() for recursive chunking of text extracted.
• Hugging Face's Sentence Transformer library for generating embeddings (model: supabase/gte-small)

  384-dimensional embeddings

• Faiss-cpu for indexing the embeddings and search/retrieval of text_chunks.

  IndexFlatL2

• Google's Gemini 1.5 Flash for text-generation
• silo is context aware. each session has a conversation history in order to preserve context (memory).

  conversations are stored as JSON objects in session_history.json, serially, marked with UUIDs [uuid4() is used]

  • session object:

    session = {
          "id" : f"{uuid}",
          "conversation" : {
                "user" : [""],
                "silo" : [""]
          }
    }