Add prompt optimization file

Author: vignesh14052002

knowledge_base/AI/PromptOptimization.md

@@ -0,0 +1,84 @@
+# Automatic Prompt Optimization
+A system that access the quality of prompt and updates it for better performance
+
+## Types
+- Without Dataset
+    - [OpenAI gpt-5 prompt optimizer](https://platform.openai.com/chat/edit?models=gpt-5&optimize=true)
+    - [Anthropic prompt improver](https://docs.claude.com/en/docs/build-with-claude/prompt-engineering/prompt-improver)
+    - [Google VertexAI prompt optimizer](https://discuss.google.dev/t/vertex-ai-prompt-optimizer-new-algorithm-sdk-for-the-general-availability/253359)
+- With Dataset (recommended)
+    - [dspy's optimizers](https://dspy.ai/learn/optimization/optimizers/)
+        - COPRO, MIPRO, SIMBA, GEPA etc
+
+## Understanding GEPA (Genetic Pareto)
+
+The goal is to improve the valset performance using the feedbacks from trainset
+
+### Terminologies
+
+Program - An entire system with one or more predictors
+Predictor - The prompt
+
+### Workflow
+- State Initialization
+    - Best programs on Valset - [{P0},{P0},{P0}]
+- While Budget available, do the following steps
+    1) Pareto Selection (Select Program)
+    2) Mini batch Selection from trainset (random sampling by default)
+    3) Generate and Evaluate Program on mini batch and collect score and feedback
+    4) Module Selection (Select Predictors to Update, round robin by default)
+    5) Make reflective dataset
+        - For each example in mini batch, frame Prompt + Generated output + evaluation feeback in context
+    6) Construct a prompt with reflective dataset as context and Use reflection model to propose new prompt
+    7) Generate and Evaluate Program on same minibatch #2 with new prompt 
+    8) Add the program with new prompt to candidate pool, if the scores are higher than #3
+    9) If Programs has more than one predictor and a common ancestor
+        - merge them (Genetic Algo) to get new program and repeat step #2 to #8
+
+### Understanding Pareto Sampling
+
+|     | E1  | E2  | E3  | Average |
+|-----|-----|-----|-----|---------|
+| P1  | 0.5 | 0.3 | 0.6 | 0.46    |
+| P2  | 0.5 | 0.4 | 0.7 | 0.53    |
+| P3  | 0.8 | 0.1 | 0.2 | 0.36    |
+
+Here If we go by max of Average score to select candidate for optimization, we would have picked P1 always
+but P3 has some information which makes it perform better on E1 will be lost
+
+So here, the candidates are P2 and P3, (leaving P1 since it is being dominated by P2 in all examples)
+from these candidates weighted sampling is used to identify the one for optimization
+in this case, P2 performs good in 2 examples, P3 performs good in 1 example, so
+```python
+import random
+random.choice(["P3","P2","P2"])
+```
+So P2 has 66% chance of being selected and P3 has 33% chance of being selected
+this is to avoid local optima
+
+### Understanding Genetic Algorithm
+[Genetic Algorithm](https://medium.com/@AnasBrital98/genetic-algorithm-explained-76dfbc5de85d) will be used to fuse best performing features from parent to create a child
+
+lets say we have a RAG system which involves two prompts (predictors)
+1) Query Generation to retrieve documents
+2) Final Answer Generation from the retieved documents
+
+lets say we have new programs P2 and P3 which are derived from P1 (common ancestor)
+we can create P4 by using predictor #1 from P2 and predictor #2 from P3
+we can create P5 by vice versa, and much more if the number of predictors under a program increases
+
+these programs will be evaluated and if there is a best performer, it will be added to the candidate pool
+
+### Things to Keep in mind
+- focus on building a robust system before trying to optimize prompt
+- quality of dataset is very important
+- feedback function should provide very detailed instructions on what went wrong
+- try your best to come up with a good base prompt
+- It is preferred to have a SOTA reflection model
+- validation dataset should be minimal and diverse
+- more iterations you run the better the results will be
+- review the optimized prompt
+
+## References
+- [Research Paper](https://arxiv.org/pdf/2507.19457)
+- [GEPA dspy Tutorial](https://youtu.be/gstt7E65FRM?si=n9QdDHgaXdMzPTQK)