SQLite Index & Performance

Chanakya MCP compiles the structural AST of your repository into an embedded SQLite database at <project>/.chanakya/index.db.

⚡

Sub-Millisecond Relational Storage

All tables are indexed on primary lookup paths: symbol names, file paths, parent IDs, and relational graph edge targets.

---

Relational Database Schema

The index is organized into eight optimized relational tables:

Table	Columns	Purpose
files	path, language, size, mtime, hash, package	Primary file registry and change-detection baseline.
symbols	id, file_path, name, kind, start_line, end_line, signature, export_kind	Core symbol declarations and location bounds.
relationships	source_id, target_id, target_name, kind, confidence	Graph edges (contains, calls, extends, etc.).
references	name_id, file_path, start_line, start_col, enclosing_symbol_id	Every semantic identifier occurrence.
reference_names	id, name	Interned dictionary mapping 2M+ reference strings to integers.
imports	file_path, specifier, imported_name, local_name, resolved_path	Static module import bindings.
exports	file_path, name, local_name, source_path, is_reexport	Exported module members and barrel bindings.
metadata	schema_version, indexed_at, total_files, total_symbols	Database version and runtime stats.

---

Reference String Interning (Schema v4)

On large repositories (e.g. 5,000+ files), storing raw strings for millions of identifier references causes database bloat.

Chanakya MCP implements reference interning:

• 2,000,000+ raw references collapse into ~55,000 distinct dictionary entries in reference_names.
• Reduces index disk footprint by over 75%, keeping memory cache hot and query times fast.

---

2-Stage Incremental Indexing

Reindexing never re-parses unchanged files:

File Modified Event
       ↓
Stage 1: Size & Mtime Check (18ms stat scan)
       ↓ (If mismatch)
Stage 2: Content SHA-256 Hash Verification
       ↓ (If hash changed)
Cascade-delete old file rows & Reinsert in single transaction

Racily-Clean Timestamp Protection

Filesystems have coarse timestamp resolution (1s on ext4, 2s on FAT). If a file is edited within two seconds of being indexed with identical byte count, Chanakya MCP does not trust the mtime alone—it evaluates the content hash to guarantee zero stale data.

---

Performance Measurements

Metric	Chanakya MCP Codebase (103 files)	Stress Monorepo (5,123 files)
Full Initial Index	1.6 seconds	~230 seconds
Incremental Reindex	18 - 24 ms	45 - 80 ms
Symbol Lookup Latency	~0.05 ms	~0.08 ms
Database Size on Disk	2.0 MB	256 MB (interned)
Extracted Symbols	655 symbols	122,936 symbols
Relational Edges	1,840 edges	370,534 relationships