SILVIA — Retrieval-Augmented Generation (RAG) System

SILVIA is a modular Retrieval-Augmented Generation (RAG) system designed to build domain-specific chatbots using PDF documents as external knowledge sources.

---

🚀 Jumpstart

To launch SILVIA:

python SILVIA.py

Follow the CLI instructions.

You can configure the RAG chatbot using:

config/parameters.yaml

By default, this configuration builds a RAG system using the PDF documents stored in:

data/Accurate

---

📁 Project Structure

config/

Configuration folder containing parameters for a test RAG named "Accurate".

---

data/

This folder contains all documents and derived artifacts used for prompt augmentation.

data/Accurate/

At the moment, it includes:

• AI Act documentation, taken from
  https://www.aiact-info.eu/full-text-and-pdf-download/

This folder will also contain:

• Chunk database

  • Currently stored as .json
  • Format example:

    [
      {
        "chunk_id": "AI_act.pdf_p1_c0",
        "text": "<chunk text>",
        "metadata": {
          "file": "AI_act.pdf",
          "page": 1
        }
      }
    ]

• FAISS index

  • faiss_index.idx
  • Stores vector embeddings for similarity search

---

Logs_and_output/

Stores outputs produced during pipeline testing.

Includes:

• Query results

  • Retrieved chunks for each query, along with similarity scores

• Augmented prompts

  • Prompts resulting from chunk retrieval and prompt assembly

• Images

  • Histograms of:
    • Chunk length distribution for the tested database
    • Query distance distribution from retrieved entries

---

ingest/

Utilities for creating a chunk index from a collection of PDF files.

Includes:

• pdf_loader.py
  Searches for all .pdf files and loads their content.

• text_cleaner.py
  Applies the following cleaning steps:

  • Replace null characters (\x00) with spaces
  • Collapse consecutive spaces and tabs into a single space
  • Reduce sequences of 3+ newlines to at most 2
  • Strip leading and trailing spaces and newlines
  • Return the cleaned text

• chunker.py
  Implements a simple chunking strategy:

  • Fixed-length chunks
  • Fixed overlap between consecutive chunks

  Chunking options (defined in the config file):

  • chunk_size
  • overlap

• build_index.py
  Manages creation of the chunk database (chunks.json)

---

embedding/

Utilities for embedding creation and vector database management.

Includes:

• build_db.py
  Creates a FAISS index:

  • data/<RAG_name>/faiss_index.idx
  • Associated metadata file

• query_db.py
  Implements the FAISSRetriever class, which:

  • Handles similarity queries
  • Can store retrieved chunks and scores during testing

---

prompting/

Utilities for assembling the final prompt provided to the language model.

---

llm/

Wrappers around Large Language Models.
Currently implemented:

• OpenAI models

---

utils/

General-purpose utility functions.

---

⚙️ Configuration File (parameters.yaml)

Name

RAG_name

Unique identifier for the RAG setup.

IMPORTANT
RAG_name must match the name of the folder inside data/ containing the documents used for prompt augmentation.

---

Ingest Options

layout:

• true: preserve original PDF layout
• false: extract text linearly

chunk_size:

Approximate number of characters per chunk.

overlap:

Number of overlapping characters between consecutive chunks.

---

Embedding and Retrieval

model:

Supported models:

• BAAI/bge-base-en-v1.5
• all-MiniLM-L6-v2

similarity:

Similarity metric:

• CosSim → cosine similarity
• L2 → Euclidean distance

retrieval_size:

Number of chunks retrieved for prompt augmentation.

---

Prompt Parameters

All prompt fields are inserted verbatim into the final prompt passed to the LLM.

tone:

Tone of the assistant’s responses (e.g. formal, friendly, natural).

verbosity:

Level of detail in responses (e.g. concise, detailed).

preambolo:

System-level instructions prepended to every query.

citation_style:

Citation format for retrieved chunks (e.g. [chunk_3]).

---

OpenAI Model Parameters

api_key:

OpenAI API key (required to use online models).

model:

LLM model (e.g. gpt-4o-mini, gpt-4.1, gpt-3.5-turbo).

temperature:

Controls randomness in generation.

max_tokens:

Maximum number of tokens in the output.

top_p:

Nucleus sampling parameter.

frequency_penalty:

Penalizes repeated tokens.

presence_penalty:

Encourages new content.

timeout:

Maximum wait time (seconds) for the API response.

---

Logging

verbose:

Enable verbose logging for debugging and monitoring.