Related work
Adjacent research improves LLM → Cypher (or general LLM decoding) along two axes Ontographia cares about: what context the model sees, and how much / how often to verify and resample. Ontographia shares the end goal (safe, accurate Cypher) but places the LLM one layer earlier. Subsetting, retries, and execution feedback therefore ship in the app / agent layer (ontographia[llm], demos under examples/llm/), not in Cypher emitters — see Architecture — outside the core.
Pipeline and design principles: Architecture.
| Work | Focus | Relation to Ontographia |
|---|---|---|
| Enhancing Text2Cypher with Schema Filtering (Ozsoy, 2025) | Prompt-time schema pruning (exact-match, NER-masked, similarity) to cut noise, hallucinations, and token cost when the model writes Cypher | Complementary for the app / agent layer: exact-match subset + full-schema fallback lives in ontographia.llm (Intent layer, issue #45). Ontographia’s core still never receives raw Cypher from the model. |
| Explore iterative refinement for Text2Cypher (Ozsoy, 2025) | Post-generation verify → correct loops over Cypher (rule-based, CyVer, execution-based, LLM-based) | Parallel concern, different object: Ontographia validates and repairs at the Intent / COM boundary, then emits Cypher deterministically. Execution→Intent refine is opt-in via examples/run_llm_e2e.py --execute --refine-on-exec (issue #46, ontographia.llm.exec_feedback). Mechanical H2 study: Evaluation — H2 (issue #55). |
| Make Every Penny Count: Difficulty-Adaptive Self-Consistency (Wang et al., NAACL 2025 Findings) | Allocate self-consistency samples by prior + posterior difficulty so easy queries avoid redundant resampling | Orthogonal cost control for Intent generation: simple intents can take one constrained decode; hard ones (large ontology, ambiguous terms) can spend more samples or a fuller schema. Does not replace ontology validation or deterministic emission. |
| Learning How Hard to Think: Input-Adaptive Allocation of LM Computation (Damani et al., ICLR 2025; arXiv:2410.04707) | Predict which inputs benefit from extra test-time compute (adaptive best-of-k, routing expensive vs cheap decoders) | Same budgeting idea for the agent layer: skip expensive Intent refinement / execution feedback when a first Intent already validates and compiles cleanly. Complements Ontographia’s cheap deterministic core rather than competing with it. |
Approach contrast
- Schema filtering & Cypher refinement assume the LLM produces Cypher and mitigate failure modes around that choice (smaller schema in the prompt; check/fix the query afterward).
- Difficulty-adaptive decoding (DSC, input-adaptive allocation) assume expensive multi-sample or multi-verifier loops and ask when to spend that compute. The Text2Cypher iterative-refinement post cites both as motivation to skip verification on simple queries.
- Ontographia keeps the LLM on schema-constrained Intent JSON, uses the ontology as the source of truth for names and structure, binds filter values as parameters, and compiles via a controlled AST. Syntax and injection risk leave the model’s job; remaining hard problems (wrong field among many plausible ones) sit in Intent generation — where schema filtering, Intent-level refinement, and difficulty-adaptive spend map naturally (app layer; see Evaluation), not Cypher rewriting.
Measured stress-test baseline (silent wrong fields still compile on large ontologies; subset prompts shrink token cost; H1 mechanical support; H2 Empty@Valid recovery mechanical support): Evaluation — Track A · H1 study · H2 study.