EXAMPLES.md

Pub/Sub Multi-Agent System - Example Configurations

Note: This deployment includes phi4-mini (general purpose), MedAIBase/MedGemma1.5:4b (medical/healthcare), and deepseek-coder:1.3b (coding) models.

Important: Use {question} in prompts to access the user's question, {schema} for database schema, and {input} for messages from subscribed topics (all case insensitive).

Example 1: Natural Language to SQL Pipeline (using phi4-mini)

Agent 1: Question Analyzer

• Title: Question Analyzer
• Prompt:

Given this database schema:
{schema}

User question: {question}

Analyze the question and identify:
1. What tables are needed
2. What columns are needed
3. What the user wants to retrieve

Respond with a clear analysis.

• Model: phi4-mini
• Subscribe Topic: START
• Publish Topic: QUESTION_ANALYZED

Agent 2: SQL Generator

• Title: SQL Generator
• Prompt:

Based on this analysis:
{input}

Database schema:
{schema}

Original question: {question}

Generate a SQL query that answers the user's question.
Return ONLY the SQL query, no explanation.

• Model: phi4-mini
• Subscribe Topic: QUESTION_ANALYZED
• Publish Topic: SQL_GENERATED

Agent 3: SQL Validator

• Title: SQL Validator
• Prompt:

Check if this SQL query is valid and correct:
{input}

Schema:
{schema}

If valid, return "VALID: " followed by the query.
If invalid, return "INVALID: " followed by the corrected query.

• Model: phi4-mini
• Subscribe Topic: SQL_GENERATED
• Publish Topic: FINAL

---

Example 2: Medical Symptom Analysis (Using MedGemma)

User Question: "Patient has fever, cough, and shortness of breath for 3 days"

Agent 1: Symptom Extractor

• Title: Symptom Extractor
• Prompt:

Patient report: {question}

Extract all symptoms mentioned in the patient report.
List them clearly and concisely.

• Model: MedAIBase/MedGemma1.5:4b
• Subscribe Topic: START
• Publish Topic: SYMPTOMS_EXTRACTED

Agent 2: Preliminary Assessment

• Title: Preliminary Assessment
• Prompt:

Based on these symptoms:
{input}

Provide a preliminary medical assessment including:
1. Possible conditions
2. Severity level
3. Recommended next steps

Note: This is for educational purposes only, not medical advice.

• Model: MedAIBase/MedGemma1.5:4b
• Subscribe Topic: SYMPTOMS_EXTRACTED
• Publish Topic: FINAL
• Subscribe Topic: SYMPTOMS_EXTRACTED
• Publish Topic: ASSESSMENT_COMPLETE

---

Example 3: Code Review and Debugging (Using DeepSeek Coder)

User Question: "Review this Python function for bugs"

Data Source:

• Label: Code
• Content:

def calculate_average(numbers):
    total = 0
    for num in numbers:
        total += num
    return total / len(numbers)

Agent 1: Bug Detector

• Title: Bug Detector
• Prompt:

Review this code for bugs:

{Code}

Identify any bugs, edge cases, or potential issues.

• Model: deepseek-coder:1.3b
• Subscribe Topic: START
• Publish Topic: BUGS_FOUND

Agent 2: Code Fixer

• Title: Code Fixer
• Prompt:

Based on these identified issues:
{input}

Original code:
{Code}

Provide the corrected code with fixes for all issues.

• Model: deepseek-coder:1.3b
• Subscribe Topic: BUGS_FOUND
• Publish Topic: FINAL

---

Example 4: Research Paper Analysis

Agent 1: Abstract Analyzer

• Title: Abstract Analyzer
• Prompt:

Paper abstract: {input}

Analyze the abstract and extract:
1. Main research question
2. Methodology
3. Key findings

• Model: phi4-mini
• Subscribe Topic: START
• Publish Topic: ABSTRACT_ANALYZED

Agent 2: Methodology Reviewer

• Title: Methodology Reviewer
• Prompt:

Based on this analysis:
{input}

Evaluate the research methodology:
1. Is it appropriate for the research question?
2. What are the strengths?
3. What are potential limitations?

• Model: phi4-mini
• Subscribe Topic: ABSTRACT_ANALYZED
• Publish Topic: METHODOLOGY_REVIEWED

Agent 3: Summary Generator

• Title: Summary Generator
• Prompt:

Based on this methodology review:
{input}

Generate a comprehensive summary of the paper suitable for a literature review.

• Model: phi4-mini
• Subscribe Topic: METHODOLOGY_REVIEWED
• Publish Topic: SUMMARY_COMPLETE

---

Example 5: Content Moderation Pipeline (using phi4-mini)

Agent 1: Content Classifier

• Title: Content Classifier
• Prompt:

Content: {input}

Classify this content into one of these categories:
- SAFE
- NEEDS_REVIEW
- INAPPROPRIATE

Provide classification and brief reasoning.

• Model: phi4-mini
• Subscribe Topic: START
• Publish Topic: CLASSIFIED

Agent 2: Detailed Review (only for NEEDS_REVIEW)

• Title: Detailed Reviewer
• Prompt:

Classification result: {input}

If the content was marked as NEEDS_REVIEW, provide:
1. Specific concerns
2. Severity level (low/medium/high)
3. Recommended action

If not NEEDS_REVIEW, just respond "NO REVIEW NEEDED"

• Model: phi4-mini
• Subscribe Topic: CLASSIFIED
• Publish Topic: REVIEW_COMPLETE

---

Example 5: Medical Symptom Analysis (phi4-mini can handle basic medical tasks)

Agent 1: Symptom Categorizer

• Title: Symptom Categorizer
• Prompt:

Patient symptoms: {input}

Categorize these symptoms by body system:
- Cardiovascular
- Respiratory  
- Neurological
- Gastrointestinal
- Other

List each symptom under its appropriate category.

• Model: phi4-mini
• Subscribe Topic: START
• Publish Topic: SYMPTOMS_CATEGORIZED

Agent 2: Diagnostic Reasoning

• Title: Diagnostic Reasoning
• Prompt:

Categorized symptoms:
{input}

Based on these symptoms, provide:
1. Most likely differential diagnoses (top 3)
2. Supporting evidence for each
3. Recommended diagnostic tests

Note: This is for educational purposes only.

• Model: phi4-mini
• Subscribe Topic: SYMPTOMS_CATEGORIZED
• Publish Topic: DIAGNOSIS_COMPLETE

---

Tips for Creating Effective Agents

1. Clear Subscribe/Publish Topics: Use descriptive names like "SQL_GENERATED" not just "STEP2"

2. Prompt Placeholders:

   • Use {question} for the user's input question
   • Use {input} for the message from the subscribed topic
   • Use {schema} to access the database schema
   • All three can be used together in any prompt

3. Model Selection:

   • phi4-mini: Versatile general-purpose model - text analysis, SQL, reasoning, classification
   • cniongolo/biomistral: Medical and scientific tasks - diagnosis, clinical reasoning, biomedical analysis

4. Final Results:

   • Publish to topic "FINAL" to display results in the Final Result box
   • Only the last agent in your chain should publish to FINAL

5. Chain Design:

   • Start simple with 2-3 agents
   • Each agent should have ONE clear purpose
   • Test each agent individually first

6. Error Handling:

   • If an agent has no subscribers for its publish topic, the chain stops
   • Make sure START has at least one subscriber
   • Check that all topics form a connected chain

7. Debugging:

   • Watch the execution log carefully
   • Check which topics have subscribers
   • Verify message flow through the bus