GPT-3.5 is not your backend database. It doesn’t matter what your company shipped last week, what your internal API looks like, or what your customer just uploaded five minutes ago. It was trained on internet-scale data, not your data.
First, let’s be real about the limits.
GPT-3.5 has a fixed brain. It doesn’t learn in real time. You can’t fine-tune it. So we rely on prompt-time context injection — feeding the model the knowledge it needs as part of the request.
This is usually done via:
- Instructive prompts
- Contextual document
- Retrieval-Augmented Generation (RAG)
Scenario: GPT-3.5, Meet My Company Wiki
Let’s say you’ve got an internal wiki. You want users to ask questions like
“What’s our refund policy for enterprise customers?”
GPT-3.5 will hallucinate like a B+ student who didn’t read the assignment — unless you give it the correct info.
Here’s how to inject that knowledge using code.
Step 1: Index Your Content
Step 2: Query-Time Retrieval
Step 3: Inject and Prompt
Now we inject that top context chunk into the GPT-3.5 prompt.
Common Pitfalls
- Context too long? GPT-3.5 has a context window limit (~4k tokens). Inject wisely.
- Chunking matters. Don’t inject whole documents. Inject the relevant slice.
- Don’t trust it blindly. Always verify generated content, especially with legal or medical info.
- Security risk: Don’t inject sensitive stuff without guardrails. The model might repeat it back where it shouldn’t.
This Is Software 2.0.
You’re not coding logic — you’re sculpting cognition. The better you get at shaping what the model sees, the smarter your app gets. No retraining is needed.
No comments:
Post a Comment