FAQ Miss
Banma SDK
2
Response
9
retrieve solution? multi-way or single
use Question + FAQ answer+ prompt to answer question
Solution 4
Performance
7
generatedocuments trunks
ASR+ NLU
Classifier
文档拆分
FAQ miss
Car ManualQuery
store embedding with document trunk ID
LLM
Vendor part
vectorizequestion
Answer
8
use document chunk IDs to retrievedocument trunks from storage
6.2.1
6.2
Which embedding models?
4
use question embedding to retrieve relevant document chunk ID
1
10
Question
Rerank
6
3
5
Local part
complex multi-intent
use Question + relevant document chunks + prompt to answer question
Vector Database
rerank document trunks
BMW Backend
voice package
Offline Part
actionxxx
Solution 1
load document
VectorDatabase
TTFT:1s-1.530 token/s70B
FAQ（Qeustion embedding: answer)
FAQ（Question embedding: answer)
Voice
6.1
11
Manual split or use tools
vectorizedocuments trunks
Banma 提供方案，部署在KBS
Customerize response
How to open sunroof?
xxxxx
Standard car manual questions
FAQ Hit
which rerank algorithm or Model?
HVD Service
KBS
4.1
No Need LLM
Solution 2
Local LLM
Arbitration
Car manual texts
FAQ hit
Open left window
BanmaBackend
Banma Backend
Speech App
Solution 3
actionXXX
embedding
Workflow
