Caching Strategy

The Caching tab allows you to connect your agent to shared memory banks. By reusing previously generated audio and text, you can achieve 0-latency responses and significantly reduce API costs.

How it Works

When the agent decides to speak or search, it first checks if it has done this exact task before.

• Scenario: Agent needs to say "Welcome to Medina Dental."
• Check: Is this phrase in the assigned Audio Cache Group?
• Hit: Yes -> Play the file instantly (0ms latency, $0 cost).
• Miss: No -> Call ElevenLabs -> Stream Audio -> Save to Cache for next time.

---

Configuration

To use caching, you must assign a Cache Group from your Data Layer to the specific function of the agent.

1. Audio Cache (TTS)

Priority: Critical.

• Function: Stores generated MP3 files.
• Strategy: Assign a group here to skip TTS generation for common phrases (Greetings, Standard Questions).
• Warning: If you change the Agent's Voice ID, you must clear or change this group, otherwise the agent will speak with mixed voices (Old Cached Voice + New Live Voice).

2. Embedding Cache (RAG)

Priority: High.

• Function: Stores vector math for text chunks.
• Strategy: Assign a group here to save money on Embedding Models (e.g., OpenAI text-embedding-3). If users ask similar questions, we reuse the math.

3. Message Cache (LLM)

Priority: Low / Experimental.

• Function: Stores the LLM's text response.
• Strategy: Only use this for Static FAQs (e.g., "Business Hours").
• Risk: Using this for dynamic conversation can lead to "Context Blindness" (The agent repeating a generic answer when a specific one was needed).

---

Auto-Caching (Learning Mode)

By default, cache is read-only. To make the agent "learn" over time, enable Auto-Cache.

• Behavior: Every time the agent generates new content (that wasn't in the cache), it saves it to the group automatically.
• Result: The first user gets standard latency. The second user gets instant responses.