Anvesh Vector DB & Knowledge Graph

Tharior Remedai uses Anvesh as its unified data engine. Anvesh provides sub-millisecond vector similarity search across indexed code repositories alongside an AST-level Knowledge Graph.

1. Knowledge Graph Grounding

Prior to LLM patch synthesis, Tharior Remedai queries Anvesh to extract:

• Exact class and function definitions (signatures, docstrings, lines).
• Caller and callee dependencies across modules.
• Related dataclasses and type definitions.

2. Vector Indexing for Remediation Patches

When an agent successfully resolves an issue and opens a pull request, the patch diff is indexed into Anvesh collection remediation_patches. If a similar issue occurs in the future, the agent retrieves the historical patch as few-shot context, improving one-shot accuracy.

3. Graph Sizing Controls & Interactive Studio Exploration

To prevent browser rendering freezing on massive monorepos with thousands of symbols, the Knowledge Graph endpoint (GET /api/v1/repos/{repo_id}/knowledge-graph) implements intelligent ego-network pruning:

• Default Limiting: Caps response payloads to 40 central nodes with guaranteed valid edge integrity.
• Query Controls: Custom bounds via ?limit=N or unconstrained full extraction via ?full=true.
• Interactive Zoom/Pan Canvas: The Web Desk provides real-time canvas zooming, panning, and node selection that links directly to the target scope in the Agent Studio.