add rag optimization documentation

README.md

@@ -166,7 +166,9 @@ To optimize your pipeline, simply define a ``Parameter`` and pass it to our ``Ge
 Whether you need to optimize task instructions or few-shot demonstrations,
 our unified framework offers an easy way to **diagnose**, **visualize**, **debug**, and **train** your pipeline.
 
-This [Trace Graph](https://adalflow.sylph.ai/tutorials/trace_graph.html) demonstrates how our auto-differentiation works.
+This [Dynamic Computation Graph](https://adalflow.sylph.ai/tutorials/trace_graph.html) demonstrates how our auto-differentiation and the dynamic computation graph work.
+
+No need to manually defined nodes and edges, AdalFlow will automatically trace the computation graph for you.
 
 ### **Trainable Task Pipeline**
 

adalflow/adalflow/datasets/hotpot_qa.py

@@ -163,12 +163,12 @@ def __len__(self):
 
 
 if __name__ == "__main__":
-    dataset = HotPotQA(root="BBH_object_counting", split="train", size=20)
+    dataset = HotPotQA(split="train", size=20)
     print(dataset[0], type(dataset[0]))
     print(len(dataset))
-    valdataset = HotPotQA(root="BBH_object_counting", split="val", size=50)
+    valdataset = HotPotQA(split="val", size=50)
     print(len(valdataset))
-    testdataset = HotPotQA(root="BBH_object_counting", split="test", size=50)
+    testdataset = HotPotQA(split="test", size=50)
     print(len(testdataset))
     print(f"valdataset[0]: {valdataset[0]}")
     print(f"testdataset[0]: {testdataset[0]}")

benchmarks/hotpot_qa/adal_exp/build_vanilla_rag.py

@@ -1,6 +1,6 @@
 """We will use dspy's retriever to keep that the same and only use our generator and optimizer"""
 
-from typing import List, Optional
+from typing import List, Optional, Union
 from dataclasses import dataclass, field
 import dspy
 
@@ -166,26 +166,26 @@ def __init__(self, passages_per_hop=3, model_client=None, model_kwargs=None):
             use_cache=True,
         )
 
+    # user should just treat it as a call function
+    # and we will handle the connection between the components
+    # they should directly pass the retriever_output along with
+    # each output's successor_map_fn.
+    # what if it is passed to two different componnents?
+    # we can create a copy
+
     def call(self, question: str, id: str = None) -> adal.GeneratorOutput:
         if self.training:
             raise ValueError(
                 "This component is not supposed to be called in training mode"
             )
-        # user should just treat it as a call function
-        # and we will handle the connection between the components
-        # they should directly pass the retriever_output along with
-        # each output's successor_map_fn.
-        # what if it is passed to two different componnents?
-        # we can create a copy
+
         retriever_out = self.retriever.call(input=question)
 
         successor_map_fn = lambda x: (  # noqa E731
             "\n\n".join(x[0].documents) if x and x[0] and x[0].documents else ""
         )
         retrieved_context = successor_map_fn(retriever_out)
 
-        # print(f"retrieved_context: {retrieved_context}")
-        # print(f"retriever_out: {retriever_out}")
         prompt_kwargs = {
             "context": retrieved_context,
             "question": question,
@@ -196,12 +196,14 @@ def call(self, question: str, id: str = None) -> adal.GeneratorOutput:
             id=id,
         )
         # self.llm.print_prompt(**prompt_kwargs)
+        # print(f"retrieved_context: {retrieved_context}")
+        # print(f"retriever_out: {retriever_out}")
         return output
 
+    # TODO: add id in the retriever output
     def forward(self, question: str, id: str = None) -> adal.Parameter:
         if not self.training:
             raise ValueError("This component is not supposed to be called in eval mode")
-        # TODO: add id in the retriever output
         retriever_out = self.retriever.forward(input=question)
         successor_map_fn = lambda x: (  # noqa E731
             "\n\n".join(x.data[0].documents)
@@ -214,6 +216,42 @@ def forward(self, question: str, id: str = None) -> adal.Parameter:
         )
         return generator_out
 
+    def bicall(
+        self, question: str, id: str = None
+    ) -> Union[adal.GeneratorOutput, adal.Parameter]:
+        """You can also combine both the forward and call in the same function.
+        Supports both training and eval mode by using __call__ for GradComponents
+        like Retriever and Generator
+        """
+        retriever_out = self.retriever(input=question)
+        if isinstance(retriever_out, adal.Parameter):
+            successor_map_fn = lambda x: (  # noqa E731
+                "\n\n".join(x.data[0].documents)
+                if x.data and x.data[0] and x.data[0].documents
+                else ""
+            )
+            retriever_out.add_successor_map_fn(
+                successor=self.llm, map_fn=successor_map_fn
+            )
+        else:
+            successor_map_fn = lambda x: (  # noqa E731
+                "\n\n".join(x[0].documents) if x and x[0] and x[0].documents else ""
+            )
+            retrieved_context = successor_map_fn(retriever_out)
+        prompt_kwargs = {
+            "context": retrieved_context,
+            "question": question,
+        }
+        output = self.llm(prompt_kwargs=prompt_kwargs, id=id)
+        return output
+
+
+def test_retriever():
+    question = "How many storeys are in the castle that David Gregory inherited?"
+    retriever = DspyRetriever(top_k=3)
+    retriever_out = retriever(input=question)
+    print(f"retriever_out: {retriever_out}")
+
 
 def test_vailla_rag():
 
@@ -248,4 +286,5 @@ def test_vailla_rag():
 
 
 if __name__ == "__main__":
-    test_vailla_rag()
+    test_retriever()
+    # test_vailla_rag()

benchmarks/hotpot_qa/adal_exp/train_vanilla.py

@@ -42,7 +42,7 @@ def __init__(
     def handle_one_task_sample(
         self, sample: HotPotQAData
     ) -> Tuple[Callable[..., Any], Dict]:
-        if self.task.training:  # TODO: make the components more clear
+        if self.task.training:
             return self.task.forward, {"question": sample.question, "id": sample.id}
         else:
             return self.task.call, {"question": sample.question, "id": sample.id}

docs/source/use_cases/index.rst

@@ -47,7 +47,9 @@ Optimization
    * - :doc:`classification`
      - Classification with `gpt-3.5-turbo`. The optimized task pipeline performs on-par with `gpt-4o`.
    * - :doc:`rag_opt`
-     - RAG and multi-hop question answering with hotpotqa dataset, two generators, and one retriever, optimizing zero-shot and few-shot learning (coming soon).
+     - Different from previous tasks where we only used one generator component, in this of hotpotqa dataset, we will demonstrates how to optimize a pipeline composed of multiple `GradComponent`(`Retriever` & `Generator`) with a standard RAG architectures
+
+.. - RAG and multi-hop question answering with hotpotqa dataset, two generators, and one retriever, optimizing zero-shot and few-shot learning (coming soon).