Appearance
Model Comparison Chat
Compare 2-4 AI models side-by-side with the same prompt. All responses stream in real-time with full PII protection.
Getting Started
Opening Comparison Mode
There are three ways to start a comparison:
- Sidebar: Click "Compare" in the left sidebar
- Starter Prompt: Hover over any starter prompt on the empty chat screen and click the comparison icon (two columns) in the top-right corner
- Carry-over from Normal Chat: Type a prompt in normal chat, then click "Compare" in the sidebar — your text carries over automatically
Selecting Models
Use the model dropdowns in the header to select 2-4 models. Click the + button to add more models (up to 4).
Features
Prompt Enhancement
The magic wand button is available in comparison mode too. Click the sparkle icon (✨) next to the submit button or press Cmd/Ctrl + Shift + E to enhance your prompt before comparing across models. This is especially useful for comparisons — a well-structured prompt highlights the real differences between models.
File Attachments
Click the paperclip icon to attach files, or drag and drop files onto the input area. Attachments are:
- Automatically scanned for PII (images go through OCR-based detection)
- Shared across all models in the comparison
- Displayed with PII badges when sensitive content is found
Large document support: Comparison mode uses a lower RAG threshold (10KB vs 1MB in regular chat). Files above 10KB are automatically indexed so only the most relevant chunks are sent to each model, reducing latency and cost when comparing with 4 models simultaneously. PII scan results from file upload are cached and reused during chat, eliminating redundant detection on large documents.
Web Search & Image Generation
Toggle Search and Image buttons below the input to enable these tools.
Tool Execution Modes
| Mode | Behavior |
|---|---|
| Per Model | Each model decides independently whether to call tools. Results may differ between models. |
| Shared | Tools run once (using the first model's decision), and results are shared as context with all models. |
Select the mode using the dropdown next to the tool buttons.
Incognito Mode
Open the ⋮ menu in the comparison header and toggle incognito mode. When active:
- Responses are not saved to the database
- The comparison won't appear in your chat history
- Performance metrics are not persisted
Keyboard shortcut: Ctrl+Shift+I
Real-Time PII Preview
As you type, GhostAI shows:
- A count of detected PII entities above the input
- An expandable list: click to see each entity with original text (truncated) and token; hover or focus an item to highlight it in the prompt (same behavior as in normal chat)
- Subject detection indicators (healthcare, legal, financial, etc.)
- This gives you visibility into what will be protected before you send
Working with Results
Model Color Coding
Each model gets a unique color (blue, purple, amber, emerald) for visual differentiation in column headers, borders, and status indicators.
Response Stats
Each completed response shows:
- TTFT (Time to First Token) — how quickly the model started responding
- Total time — end-to-end response time
- Word count — response length
Performance Waterfall
Expand the performance section below each response to see the 9-phase latency breakdown:
- PII Detection
- Tokenization
- Prompt Build
- LLM First Token
- LLM Streaming
- Tool Execution
- Detokenization
- DB Save
- Total
Voting & Promoting
- Click Vote as Winner on the best response
- Click Continue as Chat to promote the winner into a regular conversation
Feedback
Give thumbs up or down on individual model responses. Feedback is persisted to the database for analytics.
Copy & Export
- Copy: Click the copy icon on any response
- Export: Click the Export button in the header (or
Ctrl+Shift+E) to download the comparison as a Markdown file
Retry on Error
If a model fails, click Retry this model to re-run just that model without restarting the entire comparison.
Mobile Support
On small screens, model responses appear as tabs. Tap the model name tabs at the top to switch between responses.
Keyboard Shortcuts
| Shortcut | Action |
|---|---|
Enter | Send prompt |
Shift+Enter | New line |
Esc | Stop streaming |
Ctrl+Shift+E | Export comparison |
Ctrl+Shift+I | Toggle incognito |
Open the ⋮ menu in the comparison header to show/hide the keyboard shortcut hints bar.
Combined Answer (Labs) — Multi-Model Deliberation
The Combined Answer feature takes model comparison a step further by running a 3-stage deliberation pipeline that produces one clean, direct answer from multiple AI models.
How It Differs from Consensus Synthesis
| Consensus Synthesis | Combined Answer | |
|---|---|---|
| Process | Single-pass merge of initial responses | 3-stage pipeline: Initial → Deliberation → Blend |
| Deliberation | None — models don't see each other's work | Each model refines after seeing all others |
| Output | Structured analysis (verdict, agreements, contradictions) | One clean, direct answer — no analysis sections |
| Availability | Comparison chat only | Comparison chat and single chat |
Pipeline
Stage 1 — Initial Responses: Each selected model answers your prompt independently. This is the existing comparison flow — responses stream side-by-side.
Stage 2 — Deliberation: Each model receives all other models' initial responses and produces a refined answer. Models can incorporate the best ideas, correct errors, and fill gaps they see in the other responses.
Stage 3 — Blend: A synthesis model reads all the refined (deliberated) responses and merges them into one seamless, direct answer. No headers, no analysis sections — just a clean response.
Using Combined Answer in Comparison Chat
- Run a normal comparison (2-4 models)
- After all models respond, click Combined Answer (Labs badge) in the results toolbar — it appears alongside "Synthesize Consensus"
- Watch the stage progress indicator: Initial → Deliberation → Blend
- The combined result streams into a dedicated panel below the comparison columns
Actions on the Combined Result
- Copy — Copy the combined answer to your clipboard
- Export as Markdown — Download as a
.mdfile - Promote to Chat — Continue the combined answer as a regular conversation
Using Combined Mode in Single Chat
You can also use the deliberation pipeline from normal (single) chat:
- Toggle Combined Mode in the input area (look for the toggle near the send button)
- Select the models to include (2-4 models)
- Type your prompt and send — GhostAI runs the full 3-stage pipeline behind the scenes
- You receive one clean answer as the final response
Privacy & Incognito
- PII protection applies to all three stages — deliberation and blend inputs/outputs are tokenized
- Incognito mode is respected — combined answers are not saved when incognito is active
Mega (Labs) — Cross-Round Intelligence
Mega is the most powerful feature in comparison chat. It works across multiple comparison rounds, offering two modes: Mega Synthesis (structured analysis) and Mega Combine (unified answer via 3-stage deliberation).
When to Use Mega
Mega is designed for iterative research workflows:
- You ask a question, compare models, then refine and ask again
- After 2+ rounds you have a rich set of model responses, consensus results, and combined answers
- Mega reads all of them and produces a cross-round result
How It Works
- Activate: Click the Mega button (⚡ lightning icon) in the header — it appears after 2+ rounds, or after 1 history round + the current round finishing
- Select Rounds: Purple checkboxes appear on each round — including the current (live) round once all models finish. Click to select 2 or more
- Track Selection: The Mega badge shows your selection count and which data sources are available (responses, consensus, combined)
- Choose an action:
- "Synthesize N" (purple button) — structured cross-round analysis
- "Combine N" (teal button) — 3-stage deliberation pipeline producing one unified answer
- Watch It Stream: Results stream into a dedicated panel below the rounds
Mega Synthesis
Produces a structured analysis covering:
- Cross-round agreements — what all models consistently agreed on across rounds
- Evolution of understanding — how answers improved or changed between rounds
- Key insights — the most important takeaways from the full conversation arc
- Unified recommendation — a final synthesized position drawing from all rounds
Uses all available data: model responses, consensus synthesis, and combined answers from each selected round.
Mega Combine
Runs the same 3-stage deliberation pipeline as per-round Combined Answer, but across rounds:
- Initial — Each selected round's best content is loaded (combined answer > consensus > voted model > all responses)
- Deliberation — Each round's content "sees" all other rounds and produces a refined version
- Blend — A synthesis model merges all refined content into one clean, direct answer
The teal result panel shows real-time stage progress: Initial → Deliberation (2/3) → Blend
Visual Indicators
| State | What You See |
|---|---|
| Mega active | Purple "Mega" button highlighted, checkboxes on all rounds (including current) |
| Rounds selected | Purple badge with count, selected rounds get purple ring, source indicators |
| Synthesizing | Purple panel with ⚡ animation, "generating..." label |
| Combining | Teal panel with 3-stage progress dots (Initial → Deliberation → Blend) |
| Complete | Green "complete" label, copy + export buttons |
Tips
- Select non-adjacent rounds to compare how a topic evolved (e.g., rounds 1 and 5)
- The current (live) round can be included once all models finish responding
- Mega Combine picks the best content per round automatically (combined > consensus > voted > all)
- Use Synthesize for structured analysis; use Combine for a single clean answer
- Click "Clear" to deselect all rounds, or toggle the Mega button off to exit the mode
- Works with incognito mode — results are not saved when incognito is active
- Export results as Markdown with the download button
Multi-Turn Conversations
Comparison chat supports multi-turn conversations. After the first round completes, type another prompt to continue. Previous rounds are preserved as history with their votes and responses.