Skip to main content
If your agent answers from a knowledge base, through a RAG or text-to-SQL tool or knowledge inlined in its prompt, the suite needs a knowledge corpus to test it. The corpus is built from your own data and becomes the ground truth that retrieval scenarios and evals check your agent’s answers against. Without it, the suite contains no retrieval scenarios.

​
Build the corpus

Usually you don’t run anything. When your agent has a knowledge source and no corpus exists, /simulation-generate-suite asks whether to build one and builds it as part of the suite. You can also build it on its own, after the scan:
Or say it in chat: “Build a knowledge corpus from docs/help-center/.” A purely transactional agent with no knowledge source doesn’t need a corpus, and the plugin tells you so.

​
Supported sources

Databases and indexes need the SDK’s matching extraction extra, for example plurai-simulation-sdk[extract-bigquery]. The plugin installs the extras a build needs, after asking. You don’t need an LLM key. See Install the SDK.

​
What happens

1

You choose what to include

If your agent has several knowledge tools, the plugin asks which to build. For a large source it asks whether to use the whole source or a focused subset; a subset is recommended to start. If a corpus already exists, it asks whether to reuse it.
2

Your data is extracted locally

The SDK reads your source on your machine and extracts chunks and the links between them into plurai_sim/rag_corpus/. Connection details are kept in a temporary file that is deleted afterwards. Credentials never leave your machine.
3

The chunks are uploaded

Only the extracted chunks and their relationships are uploaded, as one archive. If this is your first upload, the plugin asks whether to anonymize first.
4

The backend builds the corpus

The Plurai backend builds a knowledge graph and a set of test questions from the chunks, and keeps them as ground truth. They are not written to your machine.
Each knowledge tool gets its own corpus: a personal-memory RAG and a company-docs RAG are separate knowledge bases, never merged.

​
What you get

When you build it on its own, the plugin reports that the corpus is delivered, and your next /simulation-generate-suite records it and builds the retrieval scenarios and evals from it. The suite writes the corpus references into plurai_sim/: The .knowledge_bases and .corpus_id files are committed. The build can also correct your agent specification where it disagrees with your documents. If the build comes back thin because your documents are too short or too flat, the plugin suggests rebuilding with a smaller chunk size. Rebuilding replaces the old corpus, so the plugin asks you to confirm. To remove a corpus, ask in chat: “delete this corpus.”

​
What gets uploaded

Only the extracted chunks and relationships leave your machine, never your credentials or the full source. Every upload is recorded in .plurai/uploads/audit.log. See Data and privacy.

​
Next steps

Build the suite

Generate retrieval scenarios and evals from the corpus.

Shape your simulation

Fix a wrong fact or a bad question in the corpus.

Real records

Ground scenarios in specific records instead.

Data and privacy

Anonymization, retention and deletion.