Biblios 📚

An Intelligent Semantic Discovery Engine

Biblios is a high-performance book recommendation system that transcends traditional keyword-based search. By leveraging Large Language Models (LLMs) and Vector Databases, it understands the deeper context, themes, and emotional resonance of a narrative to provide truly relevant discoveries.

---

🛠️ The Tech Stack

• Language: Python 3.13
• AI/LLM Frameworks: LangChain, Hugging Face Transformers
• Embedding Model: sentence-transformers/all-MiniLM-L6-v2 (Local Execution)
• Vector Database: ChromaDB (via langchain-chroma)
• Data Science: Pandas, NumPy, Matplotlib, Seaborn
• UI Framework: Gradio
• Secrets Management: python-dotenv

---

🚀 Project Architecture

The project is structured into five distinct engineering phases:

01. Prepare Text Data

• Exploratory Data Analysis (EDA): Performed correlation analysis using Spearman heatmaps to ensure data missingness was not biased.
• Data Cleaning: Engineered features like age_of_book and title_and_subtitle while filtering for semantic density (25-word minimum description threshold).
• Tagging: Prepended unique ISBN identifiers to descriptions to allow for precise metadata retrieval after vector matching.

02. Vector Search (The "Brain")

• Semantic Embeddings: Swapped OpenAI for a locally-hosted Hugging Face model (all-MiniLM-L6-v2) to generate high-dimensional vectors for 5,000+ books.
• Indexing: Utilized ChromaDB for efficient K-Nearest Neighbor (KNN) retrieval.
• Query Logic: Implemented a similarity search that translates natural language (e.g., "a story about redemption in the Arctic") into mathematical coordinates to find thematic neighbors.

03. Text Classification

• Zero-Shot Learning: Implemented facebook/bart-large-mnli to classify books into categories (Fiction vs. Non-Fiction) without the need for pre-labeled training data.
• Data Augmentation: Used model inference to fill gaps in the original dataset's labeling.

04. Sentiment Analysis

• Emotion Mapping: Leveraged a fine-tuned RoBERTa model to detect sentence-level sentiment across seven categories (Joy, Fear, Sadness, etc.).
• Scoring: Stored maximum probability scores to allow users to sort recommendations by their desired "vibe."

05. Gradio Dashboard

• Interactive UI: A "Glass" themed dashboard that provides a polished gallery view of book covers and metadata.
• Dynamic Controls: Users can combine semantic queries with category filters and emotional tone sorting.

---

🔧 Technical Challenges & Solutions

Building Biblios involved solving several real-world engineering obstacles:

• The Persistence Problem: Encountered duplicate entries in ChromaDB during re-indexing. Solved by implementing a manual collection wipe/reset logic before indexing runs.
• Data Parsing Issues: Fixed a ValueError during ISBN retrieval where CSV export quotes were interfering with integer conversion.
• Version Mismatches: Resolved langchain and pandas deprecation warnings (specifically the sep argument in to_csv and chunk_size constraints in CharacterTextSplitter) by upgrading to modern syntax.

---

⚙️ Installation & Setup

1. Clone the Repo:

   git clone [https://github.com/yourusername/biblios.git](https://github.com/yourusername/biblios.git)
   cd biblios

2. Install Dependencies:

   pip install pandas seaborn matplotlib langchain-huggingface langchain-chroma sentence-transformers python-dotenv gradio

3. Environment Variables: Create a .env file (see .env.example for the template). Note: This project is configured to run Hugging Face embeddings locally, reducing API dependency.

---

📊 Visuals

(Note to Self: Insert your Spearman Correlation Heatmap here to show off the EDA!)

---

Note: This project was developed as a deep dive into Semantic Search and Agentic AI workflows.