Fix support for models that do support inbuilt function calling. (#151)

* Fix: Ensure CPU is set as default device when CUDA is not used

* fix: remove redundant second LLM call

* chore: update pypdf

* fix: remove debug print statement from generate method

* fix: remove type ignore

* fix: bind tools to LLM model in agent method

* fix: formatting fixes

---------

Signed-off-by: Palaniappan R <[email protected]>

Author: palaniappan-r

backend/requirements.txt

@@ -16,7 +16,7 @@ sphinxcontrib-mermaid==0.9.2
 sphinx-book-theme==1.1.3
 sphinx-copybutton==0.5.2
 sphinx-external-toc==1.0.1
-pypdf==4.2.0
+pypdf==5.6.0
 unstructured==0.15.1
 nltk==3.9.1
 huggingface_hub[cli]==0.24.6

backend/src/agents/retriever_graph.py

@@ -133,7 +133,7 @@ def agent(self, state: AgentState) -> dict[str, list[str]]:
 
             tool_choice_chain = (
                 ChatPromptTemplate.from_template(rephrase_prompt_template)
-                | self.llm
+                | model
                 | JsonOutputParser()
             )
             response = tool_choice_chain.invoke(
@@ -143,8 +143,6 @@ def agent(self, state: AgentState) -> dict[str, list[str]]:
                 }
             )
 
-            response = model.invoke(followup_question)
-
             if response is None or response.tool_calls is None:
                 return {"tools": []}
 
@@ -189,7 +187,6 @@ def agent(self, state: AgentState) -> dict[str, list[str]]:
     def generate(self, state: AgentState) -> dict[str, Any]:
         query = state["messages"][-1].content
         context = state["context"][-1].content
-        print("state keys", state.keys())
 
         ans = self.llm_chain.invoke({"context": context, "question": query})
 

backend/src/prompts/prompt_templates.py

@@ -75,8 +75,7 @@
 Given the chat history, rephrase the follow-up question to be a standalone question.\
 The rephrased question should include only relevant information inferred from the chat history.\
 If the question is already standalone, return the same question.\
+Choose the most appropriate tools from the list of tools to answer the rephrased question.\
 Return your response as a json blob with 'rephrased_question'.\
 
-Return your response as a JSON blob with 'rephrased_question'.
-
 """

backend/src/vectorstores/faiss.py

@@ -31,7 +31,7 @@ def __init__(
     ):
         self.embeddings_model_name = embeddings_model_name
 
-        model_kwargs = {"device": "cuda"} if use_cuda else {}
+        model_kwargs = {"device": "cuda"} if use_cuda else {"device": "cpu"}
 
         self.embedding_model: Union[
             HuggingFaceEmbeddings, GoogleGenerativeAIEmbeddings, VertexAIEmbeddings