Search Settings
The Search tab controls how the assistant retrieves content from the dataset — how much context it reads and how citations appear.
Settings
| Setting | Default | What it does |
|---|---|---|
| Use full dataset contents as context | off | When on, the entire dataset is sent as context for every answer. Best for small, curated datasets (under ~100 items). |
| Max results retrieved by search engine to build response | 10 | When full-context is off, the number of matching content chunks used to build the answer. Raise it for broader answers, lower it for faster, more focused ones. |
| Show highlighted results | on | When on, the assistant shows which content items were used to build the answer — the reference cards below each response. When off, answers appear without citations. |
Choosing the Right Mode
Full context on — the assistant reads your entire dataset for every question. This guarantees it never misses a relevant item, but it's slower, uses more LLM tokens, and can become expensive on large datasets. Use it when:
- Your dataset is small (a product FAQ, a pricing page, a policy document).
- Every item is relevant to most questions.
- You need perfect recall above all else.
Full context off (the default) — the assistant searches the dataset and picks the top matches. This is faster, cheaper, and scales to thousands of items. Use it when:
- Your dataset is medium to large (a full docs site, a product catalog).
- Questions are specific and only a few items are relevant each time.
- You want to control cost.
Tuning Max Results
The default of 10 is a good starting point. From there:
| Situation | Adjust to |
|---|---|
| Answers are too narrow or miss relevant content | Raise to 15–20 |
| Answers wander or mix unrelated sources | Lower to 5–7 |
| Dataset is very small (under 50 items) | Turn on full context instead |
| Dataset is very large (5 000+ items) | Keep at 8–12 and rely on good tagging |
Show Highlighted Results
When on, each answer includes clickable reference cards — the sources the assistant used. Visitors can click through to the original page, PDF, or article. This builds trust: users can verify the answer against the source.
Turn it off only if you want a clean, chat-only experience without citations — for example, a simple FAQ bot where the answers are self-contained.
Page Preview Images
If a crawled page contains a <meta property="og:image" content="..." /> tag, the reference card for that page will include a preview image. This makes highlighted contents visually richer — visitors see a thumbnail of the source page alongside its title, summary, and URL. No extra configuration is needed: the data capture job picks up the og:image automatically during capture.
Troubleshooting
- If answers miss content you know is there, search the dataset directly to confirm the item is indexed, and re-index if needed.
- If answers are slow, you may be using full context on a large dataset — switch to search mode and tune max results.
- If reference cards show irrelevant sources, lower the max results to tighten the selection.