Combining the Best of Two Worlds: From TF-IDF to Llama LLM
summary
Practical pipeline built for thousands of product questions: deterministic rules with spaCy entity ruler to extract GitLab versions, URLs and custom “Cloud native” entities, outputting structured JSON; small embedding model and cosine clustering, then TF-IDF + embeddings to pick the top 3 representative docs per cluster; those are fed into a small Llama 2 model to generate readable summaries. Surprising finding: large fraction of tickets are integration issues (Kubernetes links), and the system directly produces prioritized doc fixes, tutorials and training material while keeping inference cost low.
AI-written summary · about
chapters (12)
- 0:00Introduction and approach
- 2:35Defining the problem
- 5:47Data analysis methodology
- 7:35Custom entity recognition
- 13:32LLM-assisted annotation
- 15:20Building the NLP pipeline
- 17:35Topic modeling and LLMs
- 23:09Prompt engineering
- 26:07Advanced pipeline extension
- 28:21High-level architecture
- 31:49Outcomes and takeaways
- 34:52Code implementation and wrap
related talks
39mCombining the Best of Two Worlds: From TF-IDF to Llama LLM
01h 30mUsing LLMs to improve your Search Engine
55mBuild enterprise generative AI apps using Llama 3 at 1,000 tokens/s on the SambaNova AI platform
42mUsing LlamaIndex for building Generative AI apps
28mThe Data That Shapes Foundational LLMs
49mBen Hall & Barbara Fusinska
29mUsing LangChain and LlamaIndex to customize LLMs...
36mBuilding & Deploying LLM Apps