What's the best way to use llama-index

kkev

What's the best way to use llama-index to retrieve row(s) and cell value from a pandas dataframe based on a natural language user query?

41 comments

LLogan M

https://docs.llamaindex.ai/en/stable/examples/query_engine/pandas_query_engine/

kkev

Thanks @Logan M . How is this different from tool use / function calling? https://discord.com/channels/1059199217496772688/1282840257800175616/1282844352929988663

LLogan M

This is just prompting the llm to write a pandas query, and then executing, and getting the LLM to interpret the result

You could certainly create a tool for an llm/agent that does the same thing

kkev

ok. any tips for improving pandas query writing based on user query? I implemented a vanilla version and the results are not good.

kkev

https://docs.llamaindex.ai/en/stable/examples/pipeline/query_pipeline_pandas/

kkev

writing the prompts from scratch?

LLogan M

Writing the prompts from scratch is probably the way to go imo. The query engine does expose hooks for your own prompts etc., but I usually like to encourage from scratch when needed

I might adapt that query pipeline to use our new workflows abstraction though, query pipelines are an older way to do this sort of thing

kkev

gotcha, i will check out workflows

kkev

can you share a link with example workflow using pandas df, if it exists or something similar?

LLogan M

Sadly, we haven't gotten that example built yet, but we have a ton of other docs and examples
https://docs.llamaindex.ai/en/stable/module_guides/workflow/#examples

LLogan M

That entire page is pretty helpful

kkev

This one looks like it could be helpful: https://docs.llamaindex.ai/en/stable/examples/workflow/sub_question_query_engine/

kkev

for working with pandas df

kkev

i am trying to figure out how workflows would work with a pandas df?

kkev

In my case, I am trying to retrieve a single value from a df column

kkev

basically, write a pandas query like df(['col1'] == 'val1' & ['col2'] == 'val2')['col3'] <-- this is what PandasQueryEngine was doing

kkev

not sure how i'd do this with workflow - do I pass the pandas df as a tool?

LLogan M

You'd have to actually execute that code, using eval()

LLogan M

or similar

kkev

Ok - the code generation itself would be done in workflows which will orchestrate prompt template, llm, response synthesis ?

kkev

@Logan M I followed the example to update the prompts: https://docs.llamaindex.ai/en/stable/examples/query_engine/pandas_query_engine/ and i am getting the error below:

Pandas Instructions:

Plain Text

df_uk[df_uk['Level 1'] == 'Business Travel'][df_uk['Level 2'] == 'Petrol car']['GHG Conversion Factor 2020']

Pandas Output: There was an error running the output as Python code. Error message: name 'df_uk' is not defined

Traceback (most recent call last):
File "/home/ec2-user/anaconda3/envs/pytorch_p310/lib/python3.10/site-packages/llama_index/experimental/query_engine/pandas/output_parser.py", line 54, in default_output_processor
output_str = str(safe_eval(module_end_str, global_vars, local_vars))
File "/home/ec2-user/anaconda3/envs/pytorch_p310/lib/python3.10/site-packages/llama_index/experimental/exec_utils.py", line 159, in safe_eval
return eval(source, _get_restricted_globals(globals), __locals)
File "<string>", line 1, in <module>
NameError: name 'df_uk' is not defined

LLogan M

is df_uk a valid variable name though? Isn't it passed in as df ?

kkev

it is passsed as a df

kkev

# Read an excel file and print all sheets
df_uk = pd.read_excel(os.getcwd() + "/data/file.xlsx", sheet_name="data")
df_uk.sample(5)

kkev

Plain Text

query_engine = PandasQueryEngine(df=df_uk, verbose=True)
prompts = query_engine.get_prompts()

kkev

Plain Text

new_prompt = PromptTemplate(
                                """\
                                You are working with a pandas dataframe in Python.
                                The name of the dataframe is `df_uk`.
                                This is the result of `print(df_uk.head())`:
                                {df_str}

                                Follow these instructions: {instruction_str}
                                
                                Query: {query_str}

                                Expression: 
                                """
    
                            ).partial_format(
                                                instruction_str = instruction_str,
                                                df_str = df_uk.head(5)
                                            )

query_engine.update_prompts({"pandas_prompt": new_prompt})

kkev

after i update the query_engine.update_prompt... it doesn't work, maybe i need to re-pass it ?

LLogan M

Seems like the llm is hallucinating the name of the df?

kkev

weird

kkev

it's included in the PromptTemplate

LLogan M

Are you using an open source llm? Not totally unexpected

kkev

nope - gpt-4o-mini

kkev

@Logan M I am using workflows , in particular RAG with Re-ranking and vectorDBs. In the linked example, https://docs.llamaindex.ai/en/stable/examples/workflow/rag/ , instead of a vectorStoreIndex with I am using MilvusVectorStore and pass the new_index in def ingest in RAGWorkflow class.

Plain Text

storage_context = StorageContext.from_defaults(vector_store=vector_store)

vector_store = MilvusVectorStore(
                                    uri="http://localhost:19530",  # set local / docker / k8s
                                    dim=384, 
                                    collection_name = collection_name,
                                    overwrite=True
                                )

storage_context = StorageContext.from_defaults(vector_store=vector_store)

new_index = VectorStoreIndex.from_documents(
                                            documents,
                                            storage_context=storage_context
                                          )

Plain Text

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[23], line 4
      1 # Run a query
      2 result = await w.run(query="What is the conversion factor for Business Travel by Diesel car in miles?", index=uk_index)
----> 4 async for chunk in result.async_response_gen():
      5     print(chunk, end="", flush=True)

AttributeError: 'VectorStoreIndex' object has no attribute 'async_response_gen'

INFO:httpx:HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
The provided information does not include details about Business Travel by Diesel car, so a conversion factor for that specific category cannot be determined from the available data.

kkev

what object do I use to iterate over or extract from <llama_index.core.indices.vector_store.base.VectorStoreIndex at 0x7fd7be5a1d50> ?

LLogan M

You can't iterate over an index 👀 you need to use a query engine and query it with aquery

kkev

this code works in the linked example using VectorStoreIndex

Plain Text

w = RAGWorkflow() 

result = await w.run(query="How was Llama2 trained?", index=index)

async for chunk in result.async_response_gen():
    print(chunk, end="", flush=True)

kkev

https://docs.llamaindex.ai/en/stable/examples/workflow/rag/

kkev

the only change is using MilvusVectorStore

kkev

what object would you call aquery on?

LLogan M

You'd attach milvus to a storage context, and use that in the index in the ingest step

Plain Text

VectorStoreIndex.from_documents(..,
., storage_context=storage_context)

LLogan M

That's the only change you'd need

Add a reply

Find answers from the community

What's the best way to use llama-index