AI Context Compiler
The AI Context Compiler transforms an exact KQP result into a bounded kristall_ai_context_bundle.
Semantic atoms
The compiler should select semantic atoms instead of independently filling unrelated arrays until a budget is exhausted.
An atom can contain enough material to interpret one unit, for example:
- assertion + subject/predicate/object + provenance;
- structural edge + endpoint identities + relation identity;
- evidence reference + digest + source/trust metadata;
- entity/property card needed to resolve referenced IDs.
This reduces the risk of including a claim while dropping the identity or provenance needed to interpret it.
Selection
Selection may consider:
- roots and required connectors;
- query intent;
- graph distance;
- property relevance;
- evidence/epistemic status;
- authority scope;
- trust class;
- byte/token budget.
Selected and omitted material remain distinguishable.
Budget model
max_bytes is the portable hard payload budget. max_tokens is optional and meaningful only with a declared tokenizer_id. Heuristic token estimates must say that they are heuristic.
Completeness
The bundle carries forward query completeness and adds compiler truncation. Semantic/evidence closure should be reported as complete, partial, not_requested or unknown.
Provenance and trust
Retrieved content retains provenance and trust class. Text included in context can inform reasoning; it does not become system policy or execution authority merely because an AI sees it.
Language
Compilation operates on semantic IDs first. Lexical resolution is a final representation layer. Missing lexicalization preserves the semantic ID rather than causing model-generated translation inside the canonical bundle.
RAG boundary
Vector/text retrieval is useful for discovery and uncrystallized source material. It does not replace exact Kristal traversal once semantic identities are known.