Background
During document retrieval, traditional RAG systems often encounter retrieval failures. Most existing approaches treat these as post-hoc problems — detecting failure after it happens and then attempting to repair it through retries or query rewriting.
However, many retrieval failures are preventable. The root cause is often not missing evidence, but a misalignment between the query and how the evidence is organized.
Proposed Approach
Vectorless will implement a built-in failure prevention system at the retrieval stage, not as a post-processing repair layer.
Core Principle: Prevent failure before it happens, rather than detect and repair after.
Key Components:
-
Semantic Tree Navigation
- Documents are organized as hierarchical semantic trees
- Agent navigates by reading node summaries, not guessing keywords
- Each step provides full visibility of all branches at current level
-
Typed Failure Awareness
- Agent recognizes different types of retrieval difficulty:
- Path not found (no relevant branch)
- Wrong branch (entered incorrect subtree)
- Insufficient evidence (found partial but incomplete)
- Different prevention strategies for different difficulty types
-
Proactive Path Adjustment
Backtrack: Return to parent node when current path is unproductiveBranchSwitch: Jump to sibling branch without going back to root- Evidence collection tracking to know when to stop
-
Bird's-Eye View
- Orchestrator sees entire corpus structure upfront
- Selects relevant documents before diving into details
- Avoids wasted exploration in irrelevant documents
Expected Benefits
| Issue |
Traditional RAG |
Vectorless Prevention |
| Similar documents causing confusion |
Happens at retrieval |
Avoided via semantic tree structure |
| Wrong branch exploration |
Discovers after, then retries |
Backtrack at next step |
| Missing cross-document evidence |
Never knows what's missing |
Orchestrator sees all top-level branches |
| Irrelevant document retrieval |
Wastes time and token |
Pre-selection before navigation |
Implementation Priority
- P0: Sub Agent with Descend/Backtrack/GetContent
- P1: BranchSwitch for same-document cross-branch navigation
- P2: Orchestrator with bird's-eye view for cross-document selection
- P3: Typed failure recognition and adaptive strategies
Non-Goals
- No persistent memory across queries
- No complex state management
- No post-retrieval repair layer (prevention, not repair)
Background
During document retrieval, traditional RAG systems often encounter retrieval failures. Most existing approaches treat these as post-hoc problems — detecting failure after it happens and then attempting to repair it through retries or query rewriting.
However, many retrieval failures are preventable. The root cause is often not missing evidence, but a misalignment between the query and how the evidence is organized.
Proposed Approach
Vectorless will implement a built-in failure prevention system at the retrieval stage, not as a post-processing repair layer.
Core Principle: Prevent failure before it happens, rather than detect and repair after.
Key Components:
Semantic Tree Navigation
Typed Failure Awareness
Proactive Path Adjustment
Backtrack: Return to parent node when current path is unproductiveBranchSwitch: Jump to sibling branch without going back to rootBird's-Eye View
Expected Benefits
Implementation Priority
Non-Goals