[Question]: Usage of Graph Algorithms and Any Slowdowns?

[nv-rliu]

Hi there,

I'm interested in understanding if haystack depends on any graph algorithms from its usage of NetworkX? If so,

• What algorithms are used for what purpose?
• What graph sizes are they being used with?
• Have users experienced any slowdowns or issues with algorithms provided by NetworkX? (Speed, algorithm availability, etc)

Furthermore, would users be interested in accelerated nx algorithms via a GPU backend? This would involve zero code change.

Any insight into this topic would be greatly appreciated! Thank you.

[lbux]

From my understanding, I think a base.py in core/pipeline takes the inputs of a haystack Pipeline and makes it into a NetworkX MultiDiGraph. That graph is then passed into the .draw() method and converted to text to be used for the Mermaid.ink server that is used for visuals. I can't speak for whether or not this is efficient as the biggest pipeline's I've created were bottlenecked by a Mermaid server issue which is now resolved.

I'm not sure if NetworkX is used elsewhere and for other purposes besides the draw() call.

[nv-rliu]

Does it then use nx.connected_components?

And otherwise, it's fair to assume that the only other use of NetworkX is just for drawing a graph.

[julian-risch]

Exactly. AzureOCRDocumentConverter uses nx.connected_components here:

haystack/haystack/components/converters/azure.py

Line 408 in 6652dd7

merged_pairs_by_page[page_idx] = [list(a) for a in list(nx.connected_components(graph))]

Drawing the pipeline graph is prepared here but it's only iterating over edges and nodes:https://github.com/deepset-ai/haystack/blob/main/haystack/core/pipeline/draw.py#L21
No advanced graph algorithms used.

[nv-rliu]

And these graph sizes are often small? <1000 nodes

[julian-risch]

Yes, <1000 nodes.

[nv-rliu]

Got it. That might be too small to see any significant speed-ups when running on GPU. But thank you for answering my questions