ExperimentIndustrial RAG
Chunking Strategies for Engineering Documents
Fixed-window chunking splits tables, tolerances and step sequences in exactly the places that carry the meaning. We tested structure-aware alternatives.
Research question
Does chunking engineering documents along their own structure — sections, steps, tables — retrieve better than fixed token windows?
Background
Work instructions, maintenance manuals and specifications are highly structured documents. A fixed token window cuts through numbered steps and parameter tables, and the retrieved fragment often loses the unit, the tolerance or the condition it applied to.
Approach
- Baseline: fixed-size chunks with overlap.
- Variant A: section-aware chunks that never split a numbered step.
- Variant B: tables extracted whole and indexed with their caption and parent section.
- Same embedding model and retrieval settings across all variants.
Observations
- Keeping tables intact removed a whole class of wrong-unit and wrong-tolerance answers.
- Section-aware chunks produced longer passages but noticeably more answerable ones.
- Overlap mattered much less once chunk boundaries followed document structure.
Limitations
- A small document set from a single equipment family.
- Reviewed qualitatively by engineers; no quantitative retrieval score was recorded.
- Document parsing quality varied by source format and was not isolated as a variable.
Next steps
- Run the same comparison on a broader document set with a recorded scoring rubric.
- Test table-aware retrieval against parameter-lookup questions specifically.
Related labs
Appeared in