Agentic Tree Search narrows RAG to three tools
Agentic Tree Search is a reference pattern and open-source implementation for agentic knowledge retrieval that keeps everything local and database-native. It models a knowledge base as a relational tree in SQL Server, then exposes just three Semantic Kernel tools to the agent: browse the map, read a node, and search nodes. The repo emphasizes zero filesystem dependence, no initial vector store, and a clean upgrade path from `LIKE` search to Full-Text or vectors later without changing the agent interface. It is demonstrated with Qwen3:8b via Ollama and positioned as a practical on-premise alternative to file-based RAG setups.
Hot take: this is less a flashy app and more a solid infra pattern for teams that already live in SQL Server and want agentic retrieval without adding a vector stack on day one.
- –The main strength is the constrained tool surface: three tools is small enough for weaker models to use reliably, which matters more than theoretical elegance.
- –The self-join tree schema preserves the folder-like mental model while keeping the storage/query layer relational and governable.
- –Starting with `LIKE` is intentionally unglamorous, but it makes the system immediately usable and easy to evolve.
- –The strongest selling point is interface stability: search and chunking can change underneath the agent without prompt or tool churn.
- –This is a good fit for enterprise/on-prem environments; it is a weaker fit if your source of truth is already a Git repo and human editorial workflow.
DISCOVERED
2h ago
2026-04-20
PUBLISHED
4h ago
2026-04-20
RELEVANCE
AUTHOR
Square-Drink3290