Kimi for Document Analysis: The Million-Token Use Case

If the question 'which clauses contradict each other across these 600 pages?' lands on your desk, a chunked retrieval system will probably miss the contradiction. The contradiction lives across documents, not inside any single chunk. A long-context model that ingests the whole corpus once can see those relationships. That is the case where Kimi earns its keep.

A reliable document-analysis prompt

The five things to set up before you ask

1. 1Strip headers, footers, and page numbers that pollute citations
2. 2Add an index page at the top: a numbered list of every document in the bundle
3. 3Mention the index explicitly in the prompt so the model uses the same labels
4. 4Repeat the task statement at the very end of the prompt — Kimi's recall is strongest at the edges
5. 5Keep your output format constraint short and explicit (table, memo, JSON)

An audit trail that survives review

Document analysis is a regulated activity in many industries. Treat every Kimi run as evidence: log the model ID, the exact prompt, the corpus hash, and the raw response. If you would not feel comfortable showing it to an auditor, do not ship it.

Apply this

• Pick a stack of 20+ related PDFs and write the index and prompt skeleton above
• Run the same prompt on a chunked RAG pipeline and on Kimi
• Diff the answers; the disagreements are where you need a human

The big idea: long context is for cross-document questions and full-corpus synthesis. For everything else, retrieval is cheaper and probably better.