Skip to main content
Retrieval is how the Agent pulls the context it needs from your knowledge to answer. A query comes in, and a small set of the most relevant chunks from the whole knowledge base is handed to the Agent. Good retrieval keeps the Agent grounded and accurate. Bad retrieval leads straight to hallucinations. Nora’s defaults are already solid, but every setting is exposed if you want to tune.
Retrieval page showing ranked results on the left and the selected chunk's full text and similarity on the right

​
The retrieval page

The Retrieval page is where you tune and test in one place.
  • Type a query in the top search bar and hit [Run]. The left column lists results by score. Pick one and the right pane shows the full chunk, neighbouring chunks, and metadata.
  • The AI polish (✦) button inside the search bar has a chat model rewrite the current query so it works better for retrieval. Requires a provider key in the workspace.
  • The page title is the active preset’s name. Click the title to open the preset menu and switch to another preset or create a new one. Double-click to rename inline.
  • The ⚙ icon on the right of the search bar opens the Retrieval settings dock. Every knob covered later on lives here.
  • The Setup / Compare toggle switches modes. Setup mode tunes one preset at a time; Compare mode runs the same query across multiple presets side by side.

​
The steps a search goes through

Each query goes through these steps before the results come back:
  1. Query: the Agent’s question. If HyDE is on, the model writes a hypothetical answer and searches with that instead.
  2. Search: the configured method (vector · keyword · hybrid · KG-rooted) gathers candidate chunks from knowledge.
  3. Ranking: each candidate’s vector and keyword scores are combined into a rank. Anything below the minimum-similarity threshold is filtered out here.
  4. Rerank (optional): if the reranker is on, the candidates are re-scored so more relevant chunks float higher.
  5. Causal reasoning (optional): if a causal graph is connected, chunks the graph knows about get boosted and related facts get pulled in.
  6. Return: the final top Top k chunks reach the Agent.
Bundle these settings together and save them as a retrieval preset.

​
Learn more

Hybrid search

Semantic similarity and keyword matching, together.

Ranking & reranking

Decide which chunks come back first.

Causal reasoning

Use the causal graph to enrich search results.

Diagnostics

Why isn’t my chunk showing up?