Add parallelization to our converters if there are noticeable performance gains

Summary

Several file converters in haystack/components/converters/ process a list of sources sequentially. For workloads with many files, adding parallelism via ThreadPoolExecutor could significantly speed up processing.

This came up in the call with @bogdankostic and @ju-gu where we noticed that Sol is using thread-based parallelism to speed up PDF processing that can occur in a query pipeline when retrieving PDF files from a web search + content fetcher request.

Motivation

Converters currently iterate over sources one at a time in their run method. Depending on the nature of the work per file, parallel execution could provide meaningful speedups:

• Network I/O-bound (highest gain): TikaDocumentConverter - requires a synchronous call to an external service. Using a thread pool executor would greatly speed this up.

• File I/O + CPU parsing (good gain): DOCXToDocument, PPTXToDocument, MSGToDocument, HTMLToDocument, XLSXToDocument, PDFToImageContent, ImageFileToImageContent - a mix of file reads (I/O, releases the GIL) and moderate CPU parsing. Threads would help primarily during the I/O phase.

• Typically CPU-bound PDF parsing (moderate gain): PyPDFToDocument, PDFMinerToDocument- parsing is pure Python and the GIL limits thread-based parallelism. ThreadPoolExecutor still helps by overlapping file reads across files, but if the pdf parsing is the dominant cost then this doesn't parallelize well with threads.

• Low priority: TextFileToDocument, MarkdownToDocument, CSVToDocument, JSONConverter, FileToFileContent - per-item work is light enough that adding the executor is probably not worth the overhead.

Investigation needed

• Benchmark sequential vs. ThreadPoolExecutor for each converter category above
• Determine appropriate default max_workers values

👋 Hello there! This issue will be handled internally and isn't open for external contributions. If you'd like to contribute, please take a look at issues labeled contributions welcome or good first issue. We'd really appreciate it!

Start in haystack/components/converters/ and compare the run methods for TikaDocumentConverter, DOCXToDocument, PPTXToDocument, MSGToDocument, HTMLToDocument, XLSXToDocument, PDFToImageContent, and ImageFileToImageContent against the lower-priority converters. Benchmark sequential processing against ThreadPoolExecutor across the listed categories, then document whether gains justify it and the appropriate default max_workers values.