Replace
whey protein
- 1pea protein
- 2hemp protein
Evaluation-backed ingredient intelligence for food product development.
A controlled LLM and retrieval system for auditable, constraint-aware product reformulation.
Try the deterministic Reformulation Copilot below. It runs in the browser with no account or API key.
Current releasereference release
- 01Parse
- 02Resolve
- 03Retrieve
- 04Constrain
- 05Rank and plan
Deterministic planning runs today. Optional model backends are candidate-locked and schema-validated.
- Validated records
- 1,700
- Engine resolution F1
- 0.9952
- Engine acceptable@3
- 1.0000
- Formulation goals
- 6
Reformulation Copilot
Paste an ingredient list, choose a formulation goal, and inspect grounded candidates from FoodLM's curated rule layer.
Candidate plan
rules groundedReplace
sugar
- 1honey
- 2maple syrup
- 3date syrup
Replace
palm oil
- 1coconut oil
- 2avocado oil
- 3high oleic sunflower oil
Replace
soy lecithin
- 1sunflower lecithin
- 2egg yolk
Replace
salt
- 1sea salt
Directional R&D suggestions only. Validate safety, sensory performance, processing, nutrition, and regulatory suitability with qualified specialists.
How it works
Evidence stays visible from label parsing through structured output.
- 01
Parse
Clean label syntax and extract ingredient phrases.
- 02
Resolve
Map phrases to a bounded ingredient ontology.
- 03
Retrieve
Union embedding neighbours with substitution priors.
- 04
Constrain
Remove candidates that violate implemented goals.
- 05
Rank and plan
Score evidence, then lock generation to candidates.
Six formulation goals, one inspectable engine
Reduce sugar
Filter full-sugar candidates before transparent ranking.
Clean a label
Compare recognised alternatives with visible tradeoffs.
Lower ingredient cost
Explore lower-cost metadata bands for review.
Remove an allergen
Exclude candidates sharing the target allergen metadata.
Improve protein
Retrieve protein alternatives and functional caveats.
Simplify a formula
Identify flagged targets and reduce label complexity.
Evaluation from committed artifacts
Verified engine results are rendered from committed artifacts. Base-model and adapted-model comparisons will be published only after attributable local runs.
| Suite | Metric | Verified result |
|---|---|---|
| Resolution | Precision | 1.0000 |
| Resolution | Recall | 0.9905 |
| Resolution | F1 | 0.9952 |
| Substitution | Acceptable@3 rate | 1.0000 |
| Substitution | Zero-unacceptable rate | 1.0000 |
| Generation integrity | JSON-schema validity | 1.0000 |
| Generation integrity | Hallucination rate | 0.0521 |
| Generation integrity | Constraint satisfaction | 0.7500 |
Engineering highlights
FoodLM separates deterministic ingredient selection from optional language generation so evidence, constraints, and private company data remain controllable.
Built for technical diligence
- Hard constraints are enforced before generation.
- Generated replacements are locked to the approved candidate pool.
- Benchmark claims render from committed evaluation artifacts.
- Private ontologies and rules plug into documented data contracts.
Open source, with the evidence boundary visible
The ontology, retrieval and ranking engine, constraint system, validated dataset, benchmark, training pipeline, and live browser demo are available for inspection.