Taking
more than ~2 minutes to decide where to transfer a call is a red flag. Even with an LLM in the loop,
routing should be milliseconds to a few seconds, not minutes. When it’s that slow, it usually means the system is trying to “search everything” (or build an enormous prompt) instead of narrowing the scope first.
A couple of common causes:
1) too much data in one bucket
If you have, say:
- 1000 extensions
- each with huge phonebooks
- and everything (including the AI Agent) in one department
…that’s like asking the agent to scan an entire library with no sections. The result is slow retrieval and expensive token usage.
Tip: Organize your PBX using
Departments so the AI Agent can filter out what it doesn’t need. Departments are your “library sections.”
2) Vector storage / knowledgebase not scoped correctly
If
Vector Storage Knowledge Base is enabled:
- How many documents did you add?
- Did you set a proper Vector Store Description?
That description is effectively
routing metadata. A good description allows the retriever/router to
exclude irrelevant stores before doing any similarity search.
If stores are not “ruled out” correctly, the system may select
all stores, then run embeddings + similarity search across all of them – which is exactly how you end up with huge delays and cost.
This description is like the routing table / ACL in a router. It must be optimized and cheap to read. This is how packets are correctly forwarded to the correct interface. We must keep network concepts in mind here.
I made an example of how a description should look like
Keep us posted on this delay.
Also where are you geographically?
