Debug Report Chunking Before LlamaIndex

A finance assistant indexes 200-page reports whose first page says every amount is in thousands.

Splitting a long report into pieces that can be read alone

A chunk is retrieved on its own and read on its own. Anything the surrounding document was carrying — the unit the figures are in, the heading that says which quarter they belong to — is gone unless the chunker puts it in the chunk. Cutting on a word count is what produces half a sentence and a page of contents at the top of the results.

Your task

Explain what each bug does to a retrieved answer, reproduce all three with the visible test, then repair chunk without changing the contract.

Pairing tasks

1. Reproduce all three failures with the visible test before editing anything.
2. Pack whole sentences up to the limit instead of cutting on a word count.
3. Carry the document note and the section heading onto every chunk.
4. Stop indexing a heading that has no body of its own.

State what the token limit costs you at each end — what a chunk too large does to precision and what a chunk too small does to the context an answer needs — and say why the overlap is measured in sentences rather than in words.

Interviewer follow-up questions

• Your chunks are clean, and a query for revenue still returns the table of contents. Why, and what do you change?
• What would you measure to choose the token limit, rather than picking one?
• When would you keep a chunk that is longer than the limit?
• The report is revised and one section is deleted. What has to happen to the index?

Skills: LlamaIndex · Chunking · Retrieval quality · Document structure · RAG indexing