Hi Guys can someone please help me with

At a glance

Hi Guys, can someone please help me with this?
When running the multiDocumentAgentDemo, how can I track the sources of the response generated by the top_agent? Here the top_agent is a agent based on the index of multiple agents

tool_mapping = SimpleToolNodeMapping.from_objects(all_tools)
obj_index = ObjectIndex.from_objects(
all_tools,
tool_mapping,
VectorStoreIndex,
)

top_agent = FnRetrieverOpenAIAgent.from_retriever(
obj_index.as_retriever(similarity_top_k=3))

response = top_agent.query("question")

11 comments

WWhiteFang_Jr

Does print(response.source_nodes) not work?

llarrygzlb

it returns a empty list []

llarrygzlb

The repsonse itself looks like this
Response(response='Boston has a thriving arts and culture scene with a variety of offerings. ... The Museum of Fine Arts, Isabella Stewart Gardner Museum, and Institute of Contemporary Art are notable institutions that house a wide range of artworks, from classical to contemporary. These venues provide a rich cultural experience for both residents and visitors.', source_nodes=[], metadata=None)

llarrygzlb

I think that the top_agent synthesize the final response based on previous response generated by the sub_agents, thus the empty list makes sense because the top_agent do not access the nodes directly... but I don't know how to check the nodes accessed by the sub_agents..

WWhiteFang_Jr

Could be! Let me check the Demo once, Let me see if I can help

llarrygzlb

Thank you very much!

WWhiteFang_Jr

Can you share the link for the demo

llarrygzlb

https://docs.llamaindex.ai/en/stable/examples/agent/multi_document_agents.html
sure

WWhiteFang_Jr

I might be wrong, but try this maybe you'll get the response sorce nodes with scores.

Plain Text

nodes_with_score = top_agent._get_tools(user_query)

I see, that in query_engine interface, Only the response is being returned.

You can also try chat, I checked the code. Source nodes are being returned in this mode.
It will be something like

Plain Text

response = top_agent.chat(user_query)

Give both of them a try. Hoping it solves your issue.

CC: @Logan M

llarrygzlb

Thank you for the information, I will give the feedback here later

llarrygzlb

@WhiteFang_Jr Thank you again for the information, they are very helpful for my question. Here is my temp solution:

I need to combine the following two steps to track the sub_agent and sources_nodes:

Step 1.

Both <top_agent.query("Query")> and <top_agent.chat("Query")> give response, choose one to generate the response.

After having the response, <top_agent.sources> will return something include the 'tool_name' indicate which one sub_agent it used.

(not clear yet whether it will return multiple 'tool_names' depend on the query in the case requires information from multiple sub_agents )

Step 2.

This requires step 1 done to have information stored in the 'engine' object

<nodes_with_score = top_agent._get_tools("Query")> returns the list of candidate sub_agents can be used

<nodes_with_score[0].query_engine.sources> get the engine object from the tool, then access to the source_nodes

With above steps, I am able to track which sub_agent(tool_name) is used, and which source_nodes the sub_agent used.

CC: @Logan M

Add a reply

Find answers from the community

Hi Guys can someone please help me with

Both <top_agent.query("Query")> and <top_agent.chat("Query")> give response, choose one to generate the response.

After having the response, <top_agent.sources> will return something include the 'tool_name' indicate which one sub_agent it used.

(not clear yet whether it will return multiple 'tool_names' depend on the query in the case requires information from multiple sub_agents )

This requires step 1 done to have information stored in the 'engine' object

<nodes_with_score = top_agent._get_tools("Query")> returns the list of candidate sub_agents can be used

<nodes_with_score[0].query_engine.sources> get the engine object from the tool, then access to the source_nodes