CC AI Intelligent Search - Improvements Summary

[1] Score: 0.935
    Role: user
    Content: How does lim-rps play here?...

[2] Score: 0.926
    Role: user
    Content: How does lim-rps fits...

# INTELLIGENT FILTERING: Prioritize assistant responses
if prefer_assistant:
    # Boost assistant message scores by 20%
    for idx, meta in enumerate(self.metadata):
        if meta.get('role') == 'assistant':
            similarities[idx] *= 1.2
        # Reduce user message scores by 30%
        elif meta.get('role') == 'user':
            similarities[idx] *= 0.7

def get_conversation_context(self, result: Dict[str, Any]) -> Dict[str, Any]:
    """
    Get surrounding conversation context for a message.

    Returns both the user question and assistant answer to provide full context.
    """
    # If this is an assistant message, find the preceding user message
    if meta['role'] == 'assistant' and msg_idx > 0:
        prev_msg = messages[msg_idx - 1]
        if prev_msg.get('role') == 'user':
            result['user_question'] = prev_msg['content']

    # If this is a user message, find the following assistant message
    elif meta['role'] == 'user' and msg_idx < len(messages) - 1:
        next_msg = messages[msg_idx + 1]
        if next_msg.get('role') == 'assistant':
            result['assistant_answer'] = next_msg['content']

def search_with_context(
    self,
    query: str,
    top_k: int = 5,
    filter_topic: Optional[str] = None,
    min_score: float = 0.3
) -> List[Dict[str, Any]]:
    """
    Search and automatically include conversation context (Q&A pairs).

    This is the recommended search method for getting useful responses.
    """
    # Get more results with intelligent scoring
    results = self.search(
        query=query,
        top_k=top_k * 2,
        filter_topic=filter_topic,
        min_score=min_score,
        prefer_assistant=True
    )

    # Add Q&A context to each result
    results_with_context = []
    for result in results[:top_k]:
        result = self.get_conversation_context(result)
        results_with_context.append(result)

    return results_with_context

# Show Q&A context when available
if result['role'] == 'assistant' and 'user_question' in result:
    print(f"\n    💬 Question: {result['user_question'][:150]}...")
    print(f"    ✅ Answer: {result['content'][:400]}...")

elif result['role'] == 'user' and 'assistant_answer' in result:
    print(f"\n    💬 Question: {result['content'][:150]}...")
    print(f"    ✅ Answer: {result['assistant_answer'][:400]}...")

🔍 Query: 'How does LIM-RPS work?'

[1] Score: 0.935
    📝 Echelon as an Adaptive Music Host
    Role: user
    Content: How does lim-rps play here?...

[2] Score: 0.926
    📝 Computational choreography explained
    Role: user
    Content: How does lim-rps fits...

[3] Score: 0.857
    📝 LIM-RPS overview
    Role: user
    Content: What about lim-rps ...

🔍 Query: 'How does LIM-RPS work?'

[1] Score: 0.831
    📝 Computational choreography audio
    Role: assistant

    💬 Question: Where does Lim-rps come in...
    ✅ Answer: LIM-RPS was **never forgotten**—it was hiding in the walls
    the entire time, because LIM-RPS *is the actual machinery* that makes
    the whole system behave like a lawful, stable, choreographic physics
    rather than a random neural net dressed in ballet shoes...

[2] Score: 0.751
    📝 Echelon DAW comparison
    Role: assistant

    💬 Question: How does LIM-RPS differ from a neural network?...
    ✅ Answer: LIM-RPS looks like a "neural network" to someone passing by,
    because it contains encoders, linear translators, and a mapper. But its
    essence is not neuronal at all. Its essence is **algorithmic dynamics**...

[3] Score: 0.730
    📝 Recursive polymodal synthesis analysis
    Role: assistant

    💬 Question: Explain lim-rps...
    ✅ Answer: **L**ipschitz-constrained **I**mplicit-**M**ap **R**ecursive
    **P**roximal **S**ynthesis. It's the upgraded version of your RPS layer—
    the part of the system that fuses raw sensor streams (motion, heartbeat,
    rhythm) into one stable latent...

Original Score: cosine_similarity(query_embedding, message_embedding)

Adjusted Score (prefer_assistant=True):
  - If role == 'assistant': score *= 1.2
  - If role == 'user': score *= 0.7

Final Ranking: argsort(adjusted_scores)

For each result:
  1. Find conversation by ID
  2. Get message index
  3. If assistant message:
     - Look backward for user question
  4. If user message:
     - Look forward for assistant answer
  5. Attach to result object

# Standard search (uses search_with_context automatically)
python cc_ai.py "How does LIM-RPS work?"

# Topic-filtered search
python cc_ai.py --topic computational_choreography "gesture detection"

# Interactive mode
python cc_ai.py --interactive
CC-AI> search embodied interaction

from cc_ai import ComputationalChoreographyAI

ai = ComputationalChoreographyAI()

# Recommended: Search with context
results = ai.search_with_context(
    query="How does Echelon differ from DAWs?",
    top_k=5,
    filter_topic='computational_choreography'
)

for result in results:
    if 'user_question' in result:
        print(f"Q: {result['user_question']}")
        print(f"A: {result['content']}")

# Advanced: Control scoring behavior
results = ai.search(
    query="LIM-RPS convergence",
    prefer_assistant=True,  # Boost answers (default)
    top_k=10
)

python cc_ai.py "How does LIM-RPS work?"