Enterprise multi-tenant
RAG platform
LLM-driven ingest · Multi-stage retrieval · Full observability
LINGXI · KB
ACME · Production · Travel KB
Employee query
"What's my expense cap for a 3-day trip to Qingdao next week?"
RAG retrieval
Rewrite · multi-recall 3 KBs
Rerank · keep Top 5
Verify · 3 valid citations
Cited answer
Qingdao is Tier-2. Lodging ¥500/night, meal ¥120/day, local transit reimbursable. Budget ≈ ¥1,860.
- [1]Travel Policy 2024.pdf · p.12
- [2]City Tier Table.xlsx · East
- [3]Meal & Transit Rules.md
100%
Citation cov.
92%
Retrieval hit
0.4s
End-to-end
№ 01 · Core capabilities
Five core capabilities
Multi-KB isolation + RBAC
Three-layer permission model: tenant / KB / folder
LLM-driven ingest
Auto structuring, chapterization, Q&A extraction
6-step RAG retrieval
Rewrite → multi-recall → rerank → verify → cite → respond
Full observability
Retrieval quality, citation accuracy, token cost at a glance
Multi-tenant architecture
Group → subsidiary → dept → individual hierarchy
№ 02 · RAG Pipeline · Live Trace
Six steps · a real query, end to end
Input: "What's my 3-day Qingdao expense cap?" — follow it through the pipeline.
"What's my expense cap for a 3-day trip to Qingdao next week?"
- STEP 01
Rewrite
Expand into 4 sub-queries: Qingdao trip / caps / tier table / meal rules
0.09s - STEP 02
Multi-recall
Vector + BM25 + keyword · 12 chunks from 3 KBs
0.18s - STEP 03
Rerank
Small-model scoring · keep Top 5
0.09s - STEP 04
Verify
Cross-check citations · drop 2 hallucinations
0.11s - STEP 05
Cite
Every fact → pdf page / xlsx cell
0.04s - STEP 06
Respond
Answer + cites + token bill + audit trail
0.03s
Qingdao = Tier-2 · Lodging ¥500/night · Meal ¥120/day · Transit reimbursable · 3-day budget ≈ ¥1,860
№ 02 · Focus & Boundaries
Our focus, and our boundaries
The greatest respect we can show a CIO is telling them upfront what we do well and what we leave alone.
What we focus on
- 01
The "last mile" of enterprise AI
We don't debate whether AI can work — we make it run inside your production systems.
- 02
Industry Skills on top of the strongest foundation models
We don't build foundation models — we make the best ones work for you.
- 03
Private knowledge & workflows into AI agents
On-prem deployment · data stays in-house · business rules stay in control.
- 04
On-site pairing + long-term operations
We don't sign off at demo — we own post-launch performance.
Our boundaries
- 01
—We don't build foundation models
We focus on industry Skills and engineering delivery — bringing the best models closer to your work.
- 02
—We don't sell generic SaaS subscriptions
We build private agents shaped to your workflow, not just a login to a generic tool.
- 03
—We don't sign off at demo
PoC is a starting line — we own runtime performance for 12 months post-launch.
- 04
—We don't take jobs we can't do well
If a scenario isn't our strength, we'll tell you — we only take on work that fits.
№ 04 · Scenarios
Where it works
№ 05 · Deployment
Deployment modes
On-premise
Data never leaves your network
Public cloud
One-click deploy on Aliyun / Tencent Cloud
Hybrid
Sensitive data private + elastic compute
№ 06 · Architecture
Technical architecture
