The complete picture, walked through side by side (4:03 → 5:53).
flowchart TB
subgraph Ingest["Ingestion side"]
DS[Data sources:<br/>docs, code, images, spreadsheets] --> RS[Restructuring layer]
RS --> SC[Structure-aware chunking]
SC --> MD[Metadata creation:<br/>summaries, keywords,<br/>hypothetical questions]
end
MD --> DB[(Database:<br/>vectors + relational)]
subgraph Query["Query side"]
UQ[User query] --> RE[Reasoning engine<br/>planner + tool execution]
RE --> MA[Multi-agent system]
end
DB <--> RE
MA --> VAL{Validation nodes:<br/>gatekeeper · auditor · strategist}
VAL --> USER[User]
VAL --> EVAL[Evaluation:<br/>qualitative · quantitative · performance]
RT[Red teaming / stress testing] -.-> Query
Ingestion side (left)
- Data sources → restructuring layer → structure-aware chunking → metadata creation (
4:21). - Metadata per chunk: summaries, extracted keywords, and hypothetical questions the chunk might answer (
4:46). - It flows into a database, not just a vector store — production usually needs embeddings and relational data working together (
4:46).
Query side
- A user question does not go straight into vector search (
5:07). - A reasoning engine with a planner figures out what the query actually needs and executes the right tools (
5:07). - A multi-agent system lets different agents work on different parts of the problem (
5:07).
Validation (bottom)
- Three "human thought" validation nodes verify the process and catch problems before anything returns to the user: gatekeeper, auditor, strategist (
5:27).
Evaluation (right)
Not "did the user click thumbs up" — actual metrics (5:27):
- Qualitative — LLM judges.
- Quantitative — precision and recall.
- Performance — latency and cost.