gamecode8

Template

Hello, I switched from QueryEngine tool to RetrieverTool for my agent, and was surprised that the text template and metadata template are overriden.

Im thinking about overriding that behavior, but just curious if there was specific reason for it that im not aware of

https://github.com/run-llama/llama_index/blob/main/llama-index-core/llama_index/core/tools/retriever_tool.py#L117

2 comments

ggamecode8

Tools

Hello, Im trying out the AgentWorkflow feature and have noticed that the tool outputs arent being captured.

To start I've made a simple single agent AgentWorkflow. While the responses are generated, the response.tool_calls list is empty, and when listening to the stream of events, I never see the ToolCallResult being output.

My goal is to be able to get the source nodes used by the query engine tool. Not sure if its an issue or I have misunderstood something. Im following https://docs.llamaindex.ai/en/stable/understanding/agent/multi_agents/

See basic example below.

topic_a_agent = FunctionAgent(
name="topic_a_expert",
description="Answers questions about topic A",
system_prompt="You are a retrieval assistant.",
tools=[QueryEngineTool((....)]
llm=OpenAI(model="gpt-4"),
)

workflow = AgentWorkflow(
agents=[topic_a_agent], root_agent="topic_a_expert"
)

response = await workflow.run(user_msg="......")

7 comments

ggamecode8

Creating a query engine tool with predefined filters

Hello, is it possible to create a query engine tool that makes use of predefined filters when retrieving?

Instead of creating multiple indexes, Im trying to use the metadata to create multiple query engine tools for a subquery engine workflow.

For example:
query engine tool 1:
use for answer with metadata topic A

query engine tool 2:
use for answer with metadata topic B

2 comments

ggamecode8

Counting tokens in llm instances without a query engine

Hello, how can i count tokens when using LLM instances in a workflow and not using a query engine as shown in the example in the docs?

Also does the TokenCountingHandler have to be set globally in a callback manager?

2 comments

ggamecode8

Handling exceptions in workflow with asyncio

Hello, what is the proper way to handle exceptions in a Workflow? In the example generator function below, even though the exception gets caught, the workflows task exception seems to still bubble up to asyncio's default exception handler. Is this expected behavior?

Plain Text

async def event_generator():
        try:
            wf = MyWorkflow(timeout=30, verbose=True)
            handler = wf.run(user_query=topic["query"])

            async for ev in handler.stream_events():
                yield {"event": "progress", "data": ev.msg}

            final_result = await handler

            # Send final result message
            yield {"event": "workflow_complete", "data": final_result}

        except Exception as e:
            error_message = f"Error in workflow: {str(e)}"
            logger.error(error_message)
            yield {"event": "error", "data": error_message}

12 comments

ggamecode8

Are workflows meant to be created on each request or reused across requests

Hello, just started trying to implement workflows and have a quick question.

In the context of a web server, are workflows meant to be created on each request or are we supposed to create one instance of a workflow and call workflow.run(…) on each request?

2 comments

ggamecode8

Docstore

Hello, can I have some clarification on what the docstore is intended to store? Chunks or full document?

The documentation here states chunks: https://docs.llamaindex.ai/en/stable/module_guides/storing/docstores/

However, I have found that the ingestion pipeline stores the full document text before chunking when using document management.

https://docs.llamaindex.ai/en/stable/module_guides/loading/ingestion_pipeline/#document-management

Thank you!

7 comments

ggamecode8

Hello is there a recommended way to

Hello is there a recommended way to rerun the ingestion pipeline in case of failure? 10K documents were inserted into the docstore but there was a failure during embedding and now rerunning it will skip them since they will be considered duplicates.

Is the solution to delete all from docstore or is there a better way?

4 comments

ggamecode8

Yes I am doing that, and works great

Yes I am doing that, and works great when I only use a single parser. But if I apply a second transformation like a text splitter, the deduping breaks in the vector store because the ref_doc_id changed after the first transformation in my code snippet.

21 comments

ggamecode8

Hello everyone! Could someone please

Hello everyone! Could someone please provide some clarification on IngestionPipeline. I am noticing that when i apply multiple transformations, the original document's ID is lost after the SentenceSplitter transformation which ends up inserting new rows into the vector store since the embedding's doc id is the doc id of the nodes from the MarkdownNodeParser transformation instead of the original document.

Is the this not the intended usage? My goal is to be able to split the markdown sections into chunks after parsing to break down long sections in my document, while preserving the original document's ID.

TIA!

Plain Text

pipeline = IngestionPipeline(
    transformations=[
        MarkdownNodeParser(),
        SentenceSplitter(chunk_size=200, chunk_overlap=0),
        OpenAIEmbedding(),
    ],
    vector_store=pg_vector_store,
    docstore=docstore
)
pipeline.run(documents=documents)

1 comment

ggamecode8

Split by HTML header | 🦜🔗 LangChain

Hello, is there a llamaindex variant of Langchains HTMLHeaderTextSplitter?

I have a tried HTMLNodeParser but the output im getting for the html i have is not great.

https://python.langchain.com/v0.1/docs/modules/data_connection/document_transformers/HTML_header_metadata/

I have tried wrapping with LangchainNodeParser but it fails because Llamaindex expects a list of strings while langchain is returning a list of document objects.

5 comments

ggamecode8

Can someone provide some tips around

Can someone provide some tips around working with 100+ documents? I have used SentenceWindowNodeParser and stored them in my vector db.

However this is performing poorly when I ask a question and expect certain sentences to be retrieved.

TIA!

10 comments

Find answers from the community

Template

Tools

Creating a query engine tool with predefined filters

Counting tokens in llm instances without a query engine

Handling exceptions in workflow with asyncio

Are workflows meant to be created on each request or reused across requests

Docstore

Hello is there a recommended way to

Yes I am doing that, and works great

Hello everyone! Could someone please

Split by HTML header | 🦜🔗 LangChain

Can someone provide some tips around