Episode 3: Technical Implementation—Fine-Tuning, Enriched Schemas, and Deployment скачать в хорошем качестве

Episode 3: Technical Implementation—Fine-Tuning, Enriched Schemas, and Deployment

On AI building for ERP series, In Episode 3, we move from the theory of RAG to the actual technical pipeline required to build a reliable ERPNext AI Agent. To cross the 90% accuracy barrier, we had to move beyond general models and implement a rigorous process of fine-tuning, data enrichment, and agentic deployment. What you will learn in this episode: • Preparing Training Data: We explain the format of our training data using "Anchors" (user questions) and "Positives" (the metadata needed to generate SQL, including tables, fields, and joins)., • Synthetic Data Generation: How we use GPT-4 mini to create "messy" or "noisy" variants of questions. This ensures the model understands real-world human phrasing like "Items with zero stock" vs. "Out of stock products.", • Enriched Schema Preparation: Moving beyond raw storage schema to an Enriched Schema. We add synonyms, descriptions, and join chunks to your Doctypes so the AI understands the "why" behind your data.,, • The Fine-Tuning Process: A walkthrough of fine-tuning the Nomic AI ModernBert embedding model using Google Colab Pro+. We explain the Loss Function—the mathematical mechanism that pushes correct results (positives) closer to the user's question in the embedding space while pushing irrelevant data away.,,, • Evaluation Metrics: How we measure success using technical benchmarks like MRR (Mean Reciprocal Rank), NDCG, and Precision to ensure the most relevant schema is always ranked at the top. • FAISS Vector Store: Building the vector store for both Master Data and Enriched Schemas to enable lightning-fast semantic search., • Deploying via Replicate & Docker: The step-by-step process of wrapping models in Docker containers (using Cog) and deploying them to Replicate for production-grade inference.,, • The Multi-Model Strategy: Why we use Qwen 4B for SQL generation and rewriting chat history, alongside a lightweight Qwen 1.5B model for formatting database results into friendly responses.,, Technical Tech Stack: • Embedding Model: Nomic AI (ModernBert) • Generative Models: Qwen 4B & 1.5B • Vector Store: FAISS • Training Environment: Google Colab Pro+ • Inference/Deployment: Replicate & Hugging Face, Join the Journey: We are building this in the open to save you from "losing a year" of failed experiments. Follow our progress as we bridge the gap between human language and ERP logic. ERPGulf - Customized ERPNext for the Gulf, regional hosting For personalized training for individuals and companies email training@ERPGulf.com https://www.ERPGulf.com   / erpgulf     / erpgulf   https://github.com/erpgulf   / erpgulf     / erpgulf     / erpgulf   https://claudion.com/