4. BSG Compression & LLM Injection

4.1 Dual-Mode Rendering

Batho Structured Graph (BSG) outputs support dual-mode rendering to align database footprint and ingestion latency with downstream use cases:

View Mode	Target Audience	Key Characteristics	Emits SYNTAX_GLUE?
storage	Downstream parsers, recovery scripts	Full-fidelity representation. Includes raw source text, byte offsets, and syntactic gaps.	Yes
agent	LLM prompts, context providers	Highly compressed representation. Includes structural definitions and signatures only.	No

View Selection Guidelines

• Storage View: Used when you need complete codebase context, cross-file references, or 100% byte-for-byte source reconstruction. It guarantees a lossless round trip.
• Agent View: Used when presenting the codebase structure to a Large Language Model (LLM). It filters out comment blocks, whitespace, and formatting anomalies, reducing token footprints by up to 10x.

4.2 Token Budget Algorithm

To prevent LLM context windows from being overwhelmed, the agent view supports token budgeting. When exporting, the engine filters and prioritizes entities using an importance-based scoring mechanism:

Figure 7: Token Budget Algorithm - Flowchart showing how the compressed agent rendering mode prioritizes entities within token constraints.

Priority Scoring Factors

Entities are scored for the agent view using the following criteria:

Factor	Weight	Description
Public API	30%	Functions, methods, and classes not prefixed with _.
Import Fan-in	25%	How many other modules reference this entity.
Semantic Tags	25%	Annotations from rule plugins (e.g. api, auth, db).
Complexity	10%	Cyclomatic complexity estimate of the AST node.
Recency	10%	Node changed in recent patch cycles.

4.3 Arrow IPC Serialization

Both storage and agent views are serialized and stored inside the .batho database. To ensure high-speed reads and minimize memory overhead when downstream tools consume these graphs:

• Arrow IPC Format: Relational data (such as entity adjacency indices and dependencies) are mapped directly to Arrow IPC table schemas, permitting memory-mapped reads without full JSON deserialization overhead.
• Binary Blobs: Compression-friendly chunks (such as individual file BSGs and relationship graphs) are compressed using zstd and stored as binary blobs in Arrow files, loaded on-demand.