AI Pipeline Architecture
LeanCore's AI pipeline uses a multi-model architecture where each stage has a specific role, just like a well-run team.
Pipeline Stages
| Stage | Role | Purpose |
|---|---|---|
| Coordinator | The Manager Plans | Deep thinking, intent analysis, retrieval planning |
| Expert | The Employee Works | Tool calling, data retrieval, answer assembly |
| Validator | The Manager Reviews | Quality gating, structured pass/fail assessment |
| Grounding | The Manager Fact-Checks | Claim extraction, evidence matching, answer repair |
How It Flows
1. Coordinator (Planning)
The Coordinator receives the user's message and:
- Analyses the complexity of the request
- Identifies which data sources and tools will be needed
- Creates a structured plan for the Expert to follow
- Determines if the task requires tool usage or can be answered from context
2. Expert (Execution)
The Expert follows the Coordinator's plan:
- Calls connected system tools to retrieve real data
- Can make multiple tool calls in sequence
- Assembles an evidence-backed response
- Has access to all tools assigned to the specialist's role
3. Validator (Quality Review)
The Validator checks the Expert's work:
- Evaluates completeness -- does it answer what was asked?
- Checks quality -- is the response well-structured and professional?
- Makes a structured pass/fail decision
- If it fails, the Expert gets another attempt
4. Grounding (Fact-Checking)
The Grounding pipeline is LeanCore's hallucination defense:
- Breaks the response into individual claims
- Verifies each claim against the actual data retrieved during execution
- Supported claims pass through unchanged
- Unsupported claims are rewritten using actual evidence
- Contradicted claims are removed entirely
Tool Routing
When a request comes in, the system automatically selects the right tools:
- The user's query is analysed for intent
- A vector search finds relevant tools from the specialist's toolkit
- Tools are ranked by relevance using routing cards (structured metadata)
- The most relevant tools are provided to the Expert
This means specialists don't need to be told which system to check -- the routing system finds the right tools automatically.
Observability
Every pipeline execution is traced end-to-end:
- Each stage is logged with timing, model usage, and token counts
- Tool calls are recorded with inputs and outputs
- Quality scores and grounding results are captured
- Full traces are available for debugging and performance analysis