Chat Modes

Hadrian's chat UI supports multiple conversation modes that define how selected models interact when responding to prompts. These modes enable sophisticated multi-model workflows beyond simple parallel responses.

Overview

Chat modes are organized into five phases based on complexity:

---

Phase 1: Core Modes

Multiple (Default)

Each model responds independently in parallel. This is the default mode.

Use cases: Getting diverse perspectives, comparing model outputs, maximum throughput.

---

Chained

Models respond sequentially, each building on previous responses.

Configuration:

Flow:

1. First model responds to original prompt
2. Each subsequent model sees all previous responses
3. Instruction: "Build upon, refine, or improve the previous response(s)"

Use cases: Iterative refinement, building on ideas, progressive deepening.

---

Routed

A router model selects which model should respond based on the prompt.

Configuration:

Flow:

1. Router model analyzes the prompt (temperature=0 for deterministic selection)
2. Router selects the best target model with reasoning
3. Selected model responds to original prompt

Use cases: Optimal model selection, cost efficiency, specialized routing.

---

Phase 2: Synthesis Modes

Synthesized

All models respond, then a synthesizer combines the results.

Configuration:

Flow:

1. Gathering phase: All non-synthesizer models respond in parallel
2. Synthesis phase: Synthesizer reads all responses and creates unified answer

Use cases: Creating definitive answers, combining expert views, resolving contradictions.

---

Refined

Models take turns improving a response through multiple rounds.

Configuration:

Flow:

1. First model generates initial response
2. Each subsequent model refines the previous response
3. Process cycles through models for N rounds (round-robin)
4. Returns final refined response

Use cases: Progressive quality improvement, error correction, iterative enhancement.

---

Critiqued

One model responds, others critique, then the original revises.

Configuration:

Flow:

1. Primary model generates initial response
2. All other models provide critiques in parallel
3. Primary model revises based on all critiques received

Use cases: Quality assurance, peer review, error detection.

---

Phase 3: Competitive Modes

Elected

Models vote democratically to select the best response.

Configuration:

Flow:

1. Responding phase: All models respond in parallel
2. Voting phase: All models vote on which response is best
3. Winner determined by vote counts

Use cases: Selecting best solution, consensus quality assurance, democratic selection.

---

Tournament

Models compete in elimination brackets with judge-based comparisons.

Configuration:

Flow:

1. Generating phase: All models respond in parallel
2. Competing phase: Models paired into brackets
3. Judge model compares each pair, selects winner
4. Winners advance to next round (bye for odd numbers)
5. Process repeats until one winner remains

Use cases: Finding best response through elimination, multi-criteria competition.

---

Consensus

Models revise their responses until they reach agreement.

Configuration:

Flow:

1. All models respond in parallel (round 0)
2. Each model sees all responses, provides revised response
3. System measures similarity/agreement between responses
4. Process repeats until threshold reached or max rounds

Use cases: Reaching agreement, finding common ground, collaborative refinement.

---

Phase 4: Advanced Modes

Debated

Models argue different positions back-and-forth, then summarize.

Configuration:

Flow:

1. Assign positions (pro/con by default) to models
2. Round 0: Each model presents opening argument from their position
3. Rounds 1-N: Each model responds to opposing arguments (rebuttal)
4. Summarizer synthesizes debate into balanced conclusion

Use cases: Exploring tradeoffs, multi-perspective analysis, arguments & counterarguments.

---

Council

Models discuss from assigned roles/perspectives, then synthesize.

Configuration:

Default roles: Technical Expert, Business Analyst, User Advocate, Risk Assessor, Innovation Specialist

Flow:

1. Assign roles (manually or auto-assigned based on query)
2. Round 0: Each model presents initial perspective from their role
3. Rounds 1-N: Each model responds to other perspectives (discussion)
4. Synthesizer combines all perspectives into comprehensive response

Use cases: Multi-stakeholder perspectives, domain expertise integration, holistic analysis.

---

Hierarchical

A coordinator decomposes tasks and delegates to worker models.

Configuration:

Flow:

1. Decomposition: Coordinator analyzes prompt, creates subtask list with assignments
2. Execution: Worker models complete assigned subtasks in parallel
3. Synthesis: Coordinator combines all worker results into final response

Use cases: Complex task breakdown, specialized workers, divide-and-conquer.

---

Phase 5: Experimental Modes

Scattershot

Run the same model multiple times with different parameter variations.

Configuration:

Default variations:

• temp=0.0 (deterministic)
• temp=0.5 (balanced)
• temp=1.0 (creative)
• temp=1.5, top_p=0.9 (very creative)

Use cases: Parameter tuning, content variation generation, creative sampling.

---

Explainer

Generate explanations at different audience levels.

Configuration:

Default levels: expert, intermediate, beginner, child, non-technical

Flow:

1. First instance explains at first audience level (e.g., expert)
2. Subsequent instances adapt/simplify for remaining levels
3. Instances cycle through levels if fewer instances than levels

Use cases: Multi-audience explanations, progressive disclosure, accessibility.

---

Confidence-Weighted

Models provide confidence scores, synthesizer weights responses accordingly.

Configuration:

Flow:

1. All non-synthesizer models respond with self-assessed confidence (0-1)
2. Responses include CONFIDENCE: [score] marker
3. Synthesizer combines responses, weighting by confidence scores

Use cases: Uncertainty-aware synthesis, reliable information prioritization.

---

Mode Selection Guide

Related Features