Auto-generated transcript. Minor errors may exist. The audio is the authoritative version.
Build Log. I'm Nick.
Here's what I shipped this week and what it taught me.
Hook
Everyone's debating fine-tuning versus RAG like it's some philosophical argument about the future of intelligence. But after shipping both to production across thirteen sites, I'll tell you something different. This isn't philosophy. It's a flowchart.
There's one question that instantly tells you which path to take. And I'm going to give you the exact framework I use to deploy both.
Context
Last month I burned four hundred dollars on GPU time because I chose wrong. I was building a customer support classifier and went straight to fine-tuning because it sounded more sophisticated. Three weeks later, I realized a simple retrieval system would have solved the problem for seven cents per day.
The cost of choosing wrong isn't academic. It's a line item on your cloud bill.
Foundation models are getting cheaper and more capable every month. Claude Haiku costs a fraction of what GPT-3.5 used to cost. This makes both techniques accessible to anyone running a real business. But it also means the stakes are higher. Get this decision right and you build something that scales. Get it wrong and you're maintaining a Frankenstein system that costs too much and breaks too often.
We're cutting through the theory. Here's what you can actually deploy this week.
The Golden Question
And this is where it gets interesting from an operations standpoint.
I've reduced the entire decision to one question: Are you asking the model to recall information, or to reason in a specific way?
Think of it like this. RAG is for looking things up in a book. Fine-tuning is for teaching the model to write a new book in your voice.
Use RAG when you need the model to pull specific, existing knowledge. Your company's internal documentation. A product catalog. Past support tickets. Customer data. The information already exists somewhere. You just need the model to find it and present it coherently.
Use fine-tuning when you need the model to output in a specific style or follow a unique reasoning pattern. Your brand's copywriting tone that took years to develop. A strict JSON schema for your API. Classification categories that are specific to your business domain.
Here's a real example from my operation. I built a support bot using RAG that answers questions based on our help documentation. When someone asks about billing cycles, it retrieves the exact section from our docs and formats a clean answer. That's recall. The information exists. The model just needs to find it.
But for the same support system, I fine-tuned Claude Haiku to classify incoming emails into twelve specific categories we defined. Things like “billing dispute with urgency flag” or “feature request with enterprise context.” Those categories don't exist in any training data. I had to teach the model our specific way of thinking about customer issues.
Recall versus reason. That's the entire decision matrix.
Start with RAG. Always. It's cheaper, faster to implement, and easier to debug. Most problems that look like they need fine-tuning can actually be solved with better retrieval and smarter prompts.
Practical Deployment
Let me walk you through the exact stacks I'm running in production. No theory. Just what works.
[BED: DUCK]
For RAG, my simplest pipeline runs like this. A webhook catches new document uploads. Python chunks them using LangChain's recursive text splitter. Embeds each chunk with OpenAI's text-embedding-3-small. Stores vectors in Pinecone with metadata tags for filtering.
When a query comes in, I retrieve the top five chunks based on cosine similarity. Inject those chunks into the context. Send the whole thing to Claude Opus for the final answer generation. This system has been running for three months with zero hallucinations on that specific data set.
[BED: SWELL]
The cost math is beautiful. That RAG system processes about two thousand queries per month. Embedding costs are negligible. Pinecone runs twelve dollars monthly. Claude API calls cost around forty dollars. Total: roughly three cents per query.
For fine-tuning, the process is surprisingly straightforward. I take five hundred to one thousand labeled examples. Format them according to the API spec. Upload to OpenAI's fine-tuning endpoint. Pay somewhere between twenty and sixty dollars depending on model size and dataset. Get back a custom model endpoint that I can call like any other API.
The key insight here is data quality over quantity. Garbage in, garbage out applies ten times more to fine-tuning.
My email classifier fine-tune cost thirty-eight dollars to train. It processes classifications at a fraction of a cent each. But I spent two weeks cleaning the training data. Making sure every example was consistent. Making sure edge cases were properly represented.
You don't need a PhD in machine learning. You need a well-documented pipeline that you can debug when things go wrong.
The Contrarian Take
You've probably heard that fine-tuning is the advanced technique.
Here's what actually happens when you run it in production.
Fine-tuning is often last-mile optimization, not a first step.
I watch teams burn six weeks collecting data and tuning models for problems that could be solved with better retrieval and smarter prompts in two days. The bias toward the sophisticated solution is real. But sophistication isn't the same as effectiveness.
A fine-tuned model is like having a pet. You have to feed it. monitor for drift when your business logic changes. Retrain when you add new product categories. It gets sick and performs poorly on edge cases you didn't anticipate.
RAG is more like cattle. If one chunk returns bad information, you fix the source document. No retraining. No model versioning. No deployment pipeline for weights.
But here's where it gets interesting. The real power move is combining them. Use RAG to pull in specific facts and current information. Use a lightly fine-tuned model to structure the response in your company's exact voice and format.
I'm running this hybrid approach for a client's customer onboarding flow. RAG retrieves relevant sections from their constantly updating product documentation. A fine-tuned model formats those facts into emails that sound exactly like their head of customer success wrote them personally.
Is your fine-tuning project a strategic necessity, or just a sexy distraction?
My bet, based on running these systems at scale, is this. For most businesses, investing in rock-solid RAG infrastructure and better prompting patterns will yield ten times better ROI than initial experiments with fine-tuning.
Mid-Roll CTA
Struggling to map this to your specific use case? I've put together a one-page decision flowchart and a production checklist for testing your RAG system. It's the exact same document I use with consulting clients.
Download it free at buildlog.dev slash rag-guide. Stop guessing. Deploy with confidence next week.
Your Homework
Here's what you're going to do today. Open a notebook. Write down your AI project's goal in one sentence. Then ask the question: Is this about recall or reason?
If it's recall, sketch out a three-step RAG pipeline. What documents or data sources? What vector database? What retrieval strategy? Keep it simple. You can always add complexity later.
If it's reason, list the first fifty examples you'd need for a fine-tuning dataset. Be specific. Don't write “good customer service response.” Write “billing dispute resolution that acknowledges frustration, provides specific next steps, and includes our standard refund policy language.”
That's it. You've just done the most valuable hour of work for your AI project. The clarity you get from this exercise will save you weeks of building the wrong thing.
Go to buildlog.dev slash tools for links to all the frameworks and platforms I mentioned today. LangChain for RAG orchestration. Pinecone for vector storage. OpenAI and Anthropic APIs for both fine-tuning and base model calls.
Cross-Promo
If you run a business and you're not using AI for the boring parts yet, that's free money on the table. I'll show you where to start. If you're building with AI in production, you'll love our sister show, Prod Deploy. We break down real-world architectures from webhooks to workflows. Search for it wherever you get podcasts.
Outro
That's the build log for this week.
Ship something. Measure it. Tell me what happened.
