# Cross-document reasoning (/docs/knowledge/retrieval/cross-document-reasoning)
Real knowledge is spread across documents that were written at different times. A 2024 policy and its 2026 update can both look like the right answer to a search. With Precision Mode on, Docklee compares every new document with the documents related to it, and records how they connect.

## How it works [#how-it-works]

1. **Summarize.** When a document is added, Docklee writes a summary and a list of the facts it states.
2. **Find related documents.** The summary is compared with the summaries of documents already in the container, and the closest ones are picked.
3. **Compare.** Each pair is compared to find where they agree, where they contradict, and whether one replaces the other.
4. **Record.** The findings are stored as connections between the documents, along with any new insight that comes from reading them together.

| Finding | Example | Effect on answers |
| --- | --- | --- |
| Connection | The pricing page explains the plans in section 3 of the contract. | Answers can draw on both. |
| Contradiction | One document says 30 days, a newer one says 45 days. | The newer document is preferred, and the conflict is known. |
| Replacement | The 2026 policy replaces the 2024 policy. | The older document stops being treated as current. |

## Limits [#limits]

| Limit | Value |
| --- | --- |
| Documents compared per new document | Up to 20, the most related ones |
| Comparisons running at once per workspace | Up to 5 |

Comparing against the most related documents, not all of them, keeps the cost of each new document the same however large the container grows.

## Availability [#availability]

Cross-document reasoning runs on containers with [Precision Mode](/docs/knowledge/retrieval/precision-mode) turned on, and uses generation credits for each document it compares. Its findings are encrypted the same way as the rest of your knowledge.

## Next steps [#next-steps]

Your documents now know about each other. See how content is split before it's indexed next.

<Cards>
  <Card title="Chunking" href="/docs/knowledge/ingestion/chunking" description="How documents are split into chunks." />
  <Card title="Precision Mode" href="/docs/knowledge/retrieval/precision-mode" description="Answers prepared at ingest." />
  <Card title="Merging and conflicts" href="/docs/okf/merging-and-conflicts" description="The same idea, for OKF concepts." />
</Cards>
