Applied GenAI: building enterprise topic discovery engines

Install Ollama in Google Colab to run llama 3.2 locally without API keys or credits, start the server, and chat with the model from Python using the Ollama package.

Install three key packages—per topic, datasets, and fasttext with its wheel and numpy 2 patch—to enable BERTopic-based topic modeling, data loading from Hugging Face, and language filtering for English text.

Implement LDA with Jensen on a New York Times corpus by cleaning text (lowercase, punctuation removal, stop words, tokenization, lemmatization), build a dictionary and document-term matrix, and discover topics.

Explore latent semantic analysis (LSA) and its algebraic approach to topic modeling, compare it with LDA, and determine the optimal number of topics using coherence scores.

Compare a purpose-built sentence transformer with RoBERTa for encoding sentence meaning. Mean pooling in RoBERTa blurs distinctions, while sentence transformers align with human judgments on SDS.

Apply singular value decomposition to sentence embeddings, compare linear SVD to nonlinear maps, and visualize semantic clustering with a locally hosted llama embedding model in 2D.

Apply k-means to 2D embeddings from UMAP to discover topics, compare default, granular, and fast configurations, and contrast centroid-based clustering with density-based approaches like HDBSCAN.

Learn to implement BERTopic for enterprise topic discovery by configuring per-topic with embedding model, a dimensionality reduction model, HDBSCAN clustering, and verbose logging, plus visualizations and semantic search.

Apply a reanker to refine topic labels with a semantic representation layer, retrain from scratch or update an existing model, and use MMR for diverse, meaningful keywords.

Build a retrieval augmented generation system with topic-based chunking guided by semantic structure, using sentence embeddings and vector stores. Leverage GPU acceleration and a device-agnostic design to handle external documents.

Build a topic-based chunking pipeline with per-topic chunking, embeddings, and a vector store, then save and load the index for no reprocessing and fast retrieval in a rag-based query flow.

Explore bonus feature after building vector store by inspecting topics with get topic model info and visualizing multilingual clusters on a 2d plane, showing IDs, sentence counts, and keywords.

Leverage LLM as a judge to evaluate open-ended answers with three independent evaluators for faithfulness, relevancy, and correctness, using a separate judge model and a vector-store based retrieval workflow.