https://github.com/praadnya/llm_rag

https://github.com/praadnya/llm_rag

Last synced: about 1 year ago
JSON representation

• Host: GitHub
• URL: https://github.com/praadnya/llm_rag
• Owner: Praadnya
• Created: 2024-01-26T11:41:56.000Z (over 2 years ago)
• Default Branch: master
• Last Pushed: 2024-01-26T12:32:03.000Z (over 2 years ago)
• Last Synced: 2025-02-24T16:32:19.210Z (over 1 year ago)
• Language: Python
• Size: 64.5 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
### RAG Pipeline:

**Components:**

- **Reader (`reader.py`):** Responsible for reading documents, extracting their content, and providing it for further processing.

- **Adapter (`adapter.py`):** Segments the document content into smaller chunks to facilitate processing and analysis.

- **Embedder (`embedder.py`):** Embeds the chunks of text into numerical vectors, typically using pre-trained language models.

- **Vector Database (`vector_database.py`):** Stores the embeddings of document chunks in a database, enabling efficient retrieval based on queries.

- **LLM(`llm.py`):** GPT4All model used with gptj as backend to generate the response.

  testpdf is an example document to run the RAG pipeline.

  ### test.py as the below flow:

  **Pipeline Execution:**

- The pipeline begins with the **`reader`** loading a document and extracting its content.

- The **`adapter`** then segments the content into chunks.

- The **`embedder`** processes each chunk to generate embeddings, representing the semantic content of the text.

- The embeddings, along with metadata, are stored in the **`vector_database`** for later retrieval.

**Querying and Retrieval:**

- A sample query (e.g., a question) is embedded using the same **`embedder`**.

- The **`vector_database`** is queried with the embedded query to retrieve the most relevant chunks.

- The retrieved chunks can serve as potential answers or relevant information.

**Language Model (LLM - Language Model):**

- The **`llm.py`** module integrates a language model (e.g., GPT-4) to generate responses based on a given context and query.

- It uses the processed document chunks as context and a user-provided query to generate responses.

**Configurations:**

- Configuration parameters such as **`chunk_size`**, **`chunk_overlap`**, and model paths are used to customize the behavior of each component.

- Embedding models and backend configurations(gptj) are specified for the language model.

**Prompt Template:**

- A **`PromptTemplate`** is used to format the input prompt for GPT-J, including placeholders for the context and query. This template aids in constructing a coherent prompt for the model.