Response styles, thinking mode, citations, file references, and streaming in Ivy.

Chat features are the controls that let you tune AI behavior, monitor context usage, and enrich responses with scientific citations in Ivy. Use them when you want shorter or more detailed answers, deeper reasoning, or references you can trace back to source databases. All of these settings live in the chat header, so you can adjust them mid-conversation without opening a separate dialog.

Chat Features

Ivy provides several controls that let you tune AI behavior, monitor context usage, and enrich responses with scientific citations. All settings are accessible from the chat header without opening a separate settings dialog.

Response styles

Response style controls how verbose and detailed the AI's output is. Click the response style chip in the chat header to switch between four presets.

Style	Description	Best for
Concise	Brief, focused answers without extended explanations.	Quick factual lookups, yes/no questions, simple tool invocations.
Balanced	Standard responses with appropriate detail. This is the default.	General research queries, exploring a topic, most day-to-day use.
Detailed	Comprehensive explanations with examples and broader context.	Learning new concepts, reviewing methodology, preparing presentations.
Technical	Precise scientific language with methodology details and literature references. Assumes domain expertise.	Writing papers, peer-review-grade analysis, expert-level discussions.

Each style adjusts the AI's temperature and appends a system prompt suffix that shapes output length and tone. Switching styles takes effect immediately on the next message.

Use Technical mode when you need the AI to cite specific literature and use precise nomenclature. The AI will assume you have domain expertise and skip introductory explanations.

Thinking mode

Thinking mode controls whether the AI performs extended reasoning before generating a response. This is especially valuable for complex multi-step queries that involve selecting and chaining multiple tools.

Mode	Token budget	Behavior
Auto	5,000 tokens	The AI decides when to use extended thinking based on query complexity. Simple lookups skip thinking; multi-tool workflows use it.
Enabled (On)	8,000 tokens	Always use extended thinking for deeper analysis, better tool selection, and more coherent multi-step workflows.
Disabled (Off)	0 tokens	Skip thinking entirely for the fastest possible responses. May reduce quality on complex queries.

Toggle between modes using the three-way switch in the chat header. The current mode is displayed as off, auto, or on.

Extended thinking consumes additional tokens from your context budget. If you are near the context limit, consider switching to Off or starting a new chat.

Context usage indicator

The context indicator in the chat header shows what percentage of the model's context window is occupied by the current conversation. It uses color coding to communicate urgency:

Color	Threshold	Meaning
Green	Below 75%	Plenty of context remaining.
Amber	75% -- 90%	Approaching the limit. Consider starting a new chat soon.
Red	Above 90%	Near the context limit. Responses may degrade or truncate.

Hover over the indicator to see exact token counts (e.g., "42,000 / 200,000 tokens"). The context limit varies by model -- Claude models support up to 200,000 tokens, while GPT-4o supports 128,000.

When the indicator turns amber, branch or start a new conversation. The AI retains no cross-chat memory, so summarize key findings before switching.

Citations

Ivy extracts citations from database and literature links in each response. The extractor recognizes URL patterns for well over a hundred sources; the most common are listed below. Citations appear as a collapsible numbered list at the bottom of each AI message.

Common databases

Database	Example identifier	URL pattern
UniProt	P68871	`uniprot.org/uniprot/P68871`
PubChem	CID 2244	`pubchem.ncbi.nlm.nih.gov/compound/2244`
NCBI Gene	672	`ncbi.nlm.nih.gov/gene/672`
PubMed	PMID 12345678	`pubmed.ncbi.nlm.nih.gov/12345678`
ChEMBL	CHEMBL25	`ebi.ac.uk/chembl/compound_report_card/CHEMBL25`
ClinicalTrials.gov	NCT01234567	`clinicaltrials.gov/study/NCT01234567`
RCSB PDB	4HJO	`rcsb.org/structure/4HJO`

How citations work

Citations are extracted through two layers:

Client-side extraction (instant) -- As the response streams in, the platform scans for markdown links matching known database URL patterns. Matches are deduplicated by source:identifier and rendered immediately.
Backend resolution (post-stream) -- After streaming completes, the backend resolves entity names (e.g., "BRCA1") to database URLs via NCBI and PubChem APIs. This runs with a 3-second timeout and is non-blocking. Resolved citations are injected into the message before it is committed to storage.

Both layers are deduplicated, so you will never see the same citation listed twice.

Clicking a citation

Click any citation number to open the source record in a new browser tab. The citation footer is collapsible -- click the header to expand or collapse the list.

File references

When the AI generates files (such as PDB structures, SDF molecules, or CSV datasets), they appear as file reference chips within the message body. Each chip shows the filename and can be clicked to preview or download.

File references are extracted from tool call results and displayed using the @filename.ext format.

SMILES preview chips

Chemical structures mentioned in SMILES notation are detected and rendered as inline preview chips. Hover over a chip to see a 2D structural diagram generated via RDKit. Click the chip to copy the SMILES string or open the compound in PubChem.

Streaming text

All AI responses stream in real time. Text updates are batched into 50 ms frames (about 20 per second). The streaming component is render-isolated, meaning the rest of the workspace does not re-render during streaming. This keeps the interface responsive even during long, detailed responses.

You can scroll, interact with canvas cards, or open action rail panels while a response is still streaming. The stream indicator in the header shows when the AI is actively generating.

Notification preferences

Per-chat notification settings are accessible from the bell icon in the chat header. Toggle individual notification channels:

Notification	Default	Description
Job completion	On	Notify when analysis jobs submitted from this chat finish running.
Mentions	On	Notify when a teammate mentions you in this chat.
Presence changes	Off	Notify when teammates join or leave the chat.
Sound	On	Play an audio alert for notifications.

Additional header actions

Action	Icon	Description
Pin/unpin	Star	Add or remove the chat from your favorites list.
Visibility	Lock/Globe	Change who can see the chat. See Collaboration.
Branch	Fork	Create a copy of the conversation to explore alternatives.
Copy link	Link	Copy a shareable URL for this chat.
Duplicate	Copy	Create a full copy of the conversation.
Clear history	Refresh	Remove all messages but keep the chat shell.
Delete	Trash	Permanently delete the conversation.

Chat Features

On this page