Evaluate OpenViking on your tasks

After your first import and retrieval, try documents or conversations from your own work. Check whether OpenViking helps you find evidence, reuse context or produce useful content. Start with one scenario, then expand after you have a result.

What you want to improve	Start here	What to check
Spend less time looking through project files	Import and retrieve documents	Find the source needed to answer the question, beyond words that merely look similar
Reuse background in a new agent session	Choose an agent integration	Confirm that memory is stored and recalled for a relevant task without entering the same background again
Organize scattered sources into readable content	Context compilation	Check which sources the output uses, whether its structure helps the reader, and what it omits or gets wrong

Check retrieval with your own sources

Choose a small set of sources you know, such as project notes, an operating manual or an incident review. Import them using the quick start. Prepare questions with known answers, recording the expected source for each. After retrieval, read the returned URI and compare it with the original text.

Include a direct-answer question, one that requires two pieces of information, and one the sources cannot answer. Record whether the evidence was found, whether anything essential was missed, and whether a subsequent agent answer follows the source. find returns URIs and scores; your agent or application still needs to read and use the content to generate an answer.

Add more sources after this first check. If results change, you can distinguish ingestion, retrieval and answer-generation problems. Use observability and diagnostics to inspect processing state and request behavior.

Check memory across two sessions

Use the capability comparison to check which capture, commit and recall steps your plugin automates. Then follow that plugin's verification instructions.

1. In the first session, provide a non-sensitive preference or piece of project context that a later task will need.
2. End or commit the session using the plugin's mechanism. Check processing state, the actual memory output and its user or Peer ownership. A successful commit does not guarantee a particular memory will be created.
3. Start a new session with the same identity. Ask for a task that needs the earlier context without entering that context again. Inspect recall records and the final answer to confirm the stored information was used.
4. Ask an unrelated question and check whether old context is carried into the answer inappropriately.

Installing a plugin is only part of verifying memory. If no memory was produced, inspect sessions and extraction. If memory exists but is not recalled, inspect identity, retrieval scope and plugin behavior. For applications that manage their own sessions, follow the application workflows to the relevant APIs.

Check one compiled output

Choose a knowledge wiki, knowledge graph or daily report. Follow the tutorial to prepare sources and run the task. Confirm completion and the output location, then review the content:

• Can you verify key facts against their sources?
• What is missing, repeated or unsupported?
• Can you use the output for the reading or organizing task you intended?

Start with familiar sources so you can recognize mistakes. Task completion and generated files show that processing finished; the output still needs review before you rely on it.

Choose what to do next

Your result	Next step
Retrieval meets your needs and you want it in a daily tool	Connect an existing agent
You have verified the capability and want it in your application	Application workflows
You want to use the agent included with OpenViking	VikingBot Installation and Configuration
You are ready to share a service with your team	Choose a deployment path, then verify identity and access control
You want to understand why the system behaves this way	Concepts and principles

Keep your questions, expected sources and observed results. Reuse them when changing configuration or upgrading. When comparing results before and after a change, keep tasks, models and inputs consistent; a retrieval score or one successful run does not establish an improvement.