Decision Framework
Khi nào dùng gì? Framework quyết định:
1. Prompt Engineering
Dùng khi:
- Task đơn giản, có thể describe trong prompt
- Knowledge đã có sẵn trong model
- Cần iterate nhanh
Ví dụ: Summarization, translation, simple Q&A
2. RAG
Dùng khi:
- Knowledge thường xuyên update
- Cần cite sources
- Domain knowledge không có trong base model
- Privacy concerns (data stays in your database)
Ví dụ: Customer support với product docs, legal Q&A, internal wiki search
3. Fine-tuning
Dùng khi:
- Cần consistent style/format
- Task phức tạp với nhiều examples
- Latency critical (RAG adds retrieval time)
- Cost critical (long prompts = expensive)
Ví dụ: Code completion theo company style, domain-specific terminology, output format standardization
Comparison Matrix
| Criteria | Prompt | RAG | Fine-tune |
|---|---|---|---|
| Setup time | Minutes | Hours | Days |
| Data needed | 0 | Documents | 100-10K examples |
| Update frequency | Instant | Hours | Days |
| Latency | Low | Medium | Low |
| Cost per query | Medium | Medium | Low |
| Accuracy ceiling | Medium | High | Very High |
Hybrid Approaches
RAG + Fine-tuning: Fine-tune model cho format/style, RAG cho knowledge Fine-tune + Prompt: Fine-tune base capability, prompt cho task-specific instructions
Decision Flowchart
Có cần external knowledge?
→ Yes: RAG
→ No: Có cần consistent format/style?
→ Yes: Fine-tuning
→ No: Prompt Engineering
Pro Tips
- Start simple: Prompt → RAG → Fine-tuning (tăng dần complexity)
- Combine: RAG + Fine-tuned model often beats either alone
- Evaluate first: Benchmark current approach trước khi upgrade
🔥 80% use cases có thể solve với RAG. Fine-tuning cho 20% cần extreme customization.
