Kimi for Document Analysis: The Million-Token Use Case

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-moonshot-document-analysis-creators

1. A business analyst needs to find whether any clauses in a 600-page regulatory filing contradict each other. Which approach would likely yield the best result?

   1. Split the documents into 50-page chunks and query each separately
   2. Use a chunked RAG system with semantic search across the document
   3. Feed the entire corpus to a long-context model with a contradiction-detection prompt
   4. Use a vector database with keyword matching

2. What is the primary weakness of chunked retrieval systems when analyzing large document corpora?

   1. They may miss information that spans multiple documents or chunks
   2. They always produce inaccurate citations
   3. They cannot handle more than 10,000 tokens at once
   4. They cannot summarize content effectively

3. Before running a document analysis prompt on Kimi, why should you strip headers, footers, and page numbers from the input documents?

   1. They trigger safety filters in the model
   2. They can pollute citations and cause the model to reference page numbers instead of content
   3. The model cannot read numbers in any format
   4. They consume too many tokens and increase costs

4. What is the purpose of adding a numbered index of documents at the top of a Kimi prompt?

   1. It helps the model understand the total token count of the corpus
   2. It provides consistent labeling that the model can reference in citations
   3. It enables the model to access the documents without reading them
   4. It reduces the processing time for large document bundles

5. Why does the lesson recommend repeating the task statement at the very end of a Kimi prompt?

   1. It reduces the total token count needed
   2. It makes the prompt grammatically correct
   3. It prevents the model from generating harmful outputs
   4. Kimi's recall is strongest at the edges of a prompt

6. What does the 'OCR ceiling' refer to in document analysis?

   1. The cost threshold where long-context becomes unfeasible
   2. The degradation of output quality when inputs are scanned images without extractable text
   3. The maximum number of pages a model can process
   4. The limit on how many documents can be analyzed at once

7. What information should be logged for an audit trail when using Kimi for regulated document analysis?

   1. The number of API calls made and total cost incurred
   2. The model author's personal information and company address
   3. The model ID, exact prompt used, corpus hash, and raw response
   4. The names of all employees who viewed the analysis

8. In the context of document analysis, what does 'grounding' refer to?

   1. The model's ability to generate creative new content
   2. Reducing the token count of the input corpus
   3. The physical location where document processing occurs
   4. Tying the model's outputs to specific citations from the input documents

9. What happens when poor-quality scanned PDFs are fed into a long-context document analysis system?

   1. The model will refuse to process the documents
   2. The model automatically corrects OCR errors in its output
   3. The model will faithfully amplify the errors at scale
   4. The system will synthesize coherent summaries regardless of input quality

10. What is a key cost disadvantage of using Kimi for simple single-fact lookups compared to chunked RAG?

    1. Kimi requires human reviewers for every query
    2. Chunked RAG cannot perform fact lookups
    3. Chunked RAG has lower cost per query for simple lookups
    4. Kimi charges based on document length only

11. Based on the lesson's comparison table, which task is chunked RAG particularly good at?

    1. Multi-document synthesis
    2. Single fact lookup
    3. Cross-document contradiction detection
    4. Full-corpus summarization

12. What output format constraint does the lesson recommend for document analysis prompts?

    1. Allow the model to choose any format it prefers
    2. Require XML output with custom tags
    3. Keep it short and explicit—use table, memo, or JSON formats
    4. Use lengthy narrative paragraphs with detailed explanations

13. What does the lesson identify as the core use case where long-context document analysis earns its premium cost?

    1. When you need to look up a single phone number from a contract
    2. When the documents contain mostly images rather than text
    3. When answering questions that require synthesizing information across many documents
    4. When you need the cheapest possible solution

14. A user reports dropping an entire codebase into Kimi and getting a coherent summary. Why might chunked RAG struggle with the same task?

    1. Chunked RAG is faster but produces identical results
    2. Cross-file dependencies and patterns may span chunks that retrieval misses
    3. Codebase summaries require images that chunked systems cannot process
    4. Chunked RAG cannot read code files

15. What should you do if your document corpus contains many scans of scans (poor quality images)?

    1. Use a video format instead of text
    2. Split them into smaller files
    3. Feed them directly to Kimi anyway—it will figure it out
    4. Clean the inputs with proper OCR before trusting any synthesis