Understand the Components of a RAG System

A RAG system combines:

• Retrieval: Fetching relevant data from a database or knowledge base.
• Generation: Using a language model to generate a natural language response based on the retrieved data.

The hope for this project is to:

1. Parse the user's query.
2. Translate it into a database query (e.g., SQL).
3. Execute the query to retrieve the data.
4. Generate a response based on the retrieved data.

System Design

a. Natural Language Understanding (NLU)

• Use a language model (TBD) to understand the user's query.
• Extract key information from the query, such as:
  • Metric: "How many people"
  • Condition: "spent more than $200"
  • Timeframe: "past 10 days"
  • ETC

b. Query Translation

• Convert the parsed query into a database query (e.g., SQL).
  • Example: "How many people spent more than $200 in the past 10 days" SQL:

SELECT COUNT(DISTINCT user_id)
FROM transactions
WHERE amount > 200
  AND transaction_date >= NOW() - INTERVAL '10 days';

c. Database Interaction

• Connect to your database (e.g., PostgreSQL, MySQL, etc.).
• Execute the generated query and retrieve the results.

d. Response Generation

• TBD on how the response should be. The response can be a simple count, a visualization, or a natural language sentence.

5. Challenges and Improvements

• Query Translation Accuracy
• Error Handling
• Performance
• Security

To help the model to understand the data in the database, we will feed it with the schema information. Example: table names, column names, data types, relationships, etc.

{
  "tables": [
    {
      "name": "transactions",
      "columns": [
        {
          "name": "user_id",
          "type": "integer",
          "foreign_key": "users.id"
        },
        {
          "name": "amount",
          "type": "float"
        },
        {
          "name": "transaction_date",
          "type": "date"
        }
      ]
    },
    {
      "name": "users",
      "columns": [
        {
          "name": "id",
          "type": "integer",
          "primary_key": true
        },
        {
          "name": "name",
          "type": "text"
        }
      ]
    }
  ]
}

This process will be automated whenever a user tries to interact with the system. The system will automatically fetch the schema information from the database and feed it to the language model.

Problems to resolve

1. Passing the schema information to the model causes an error, and so we need to find a better way of doing it.

Token indices sequence length is longer than the specified maximum sequence length for this model (2504 > 512). Running this sequence through the model will result in indexing errors