OpenAI GPT-5.6 Sol
Used for structured data extraction and graph resolution.
Dairy products* - Milk, cheese, and yogurt are primary sources.
Meat and poultry* - Beef, lamb, pork, and chicken provide significant amounts.
Fish* - Salmon, tuna, and mackerel.
Eggs* - Specifically the egg yolk.
Green vegetables* - Spinach, broccoli, asparagus, and avocado.
Whole grains and fortified foods* - Whole grain breads, cereals, and wheat germ.
Legumes* - Soybeans, almonds, and various nuts.
Animal-derived foods* - Exclusively found in significant amounts in animal tissues; plant foods generally lack B12 unless fortified.
Organ meats* - Liver and kidney (beef, lamb) are exceptionally high in B12.
Seafood and fish* - Clams, oysters, mussels, caviar, octopus, sardines, salmon, tuna, and trout.
Meat and poultry* - Beef, lamb, pork, chicken, and turkey.
Dairy products* - Milk, yogurt, and cheese.
Eggs* - Whole eggs.
Fortified foods* - Fortified breakfast cereals, nutritional yeasts, and plant-based milk substitutes (soy, almond, oat) for those following a strict vegetarian or vegan diet.
Bioavailability* - The absorption of Vitamin B12 depends heavily on adequate stomach acid and intrinsic factor production.
Cooking losses* - Both B2 and B12 can be lost in cooking water or destroyed by excessive heat and prolonged light exposure; gentle cooking methods are preferred.
Vitamin B2 ->
-> Vitamin B2
Vitamin B12 ->
-> Vitamin B12
Food ->
-> Food
A source-grounded pipeline turns Acu-Cell pages into a map of entities and claims. The graph organizes what the source says; it does not invent a medical answer or certify that a claim is correct.
The source pages are converted into clean, ordered text. Section context is retained so a sentence is interpreted with the heading and surrounding material that give it meaning.
A schema-constrained language model reads one complete source record at a time and identifies entities, directed relationships, endpoint roles, qualifiers and a rationale. The extraction rules prohibit creating a relationship from proximity or formatting alone.
Dose, timing, certainty, evidence and other conditions remain qualifiers on the precise relationship they modify. General facts about an entity become attributes. This prevents a qualified statement from being displayed as an unconditional one.
Name similarity, shared attributes, explicit identity statements and mutual semantic similarity produce 784 candidate groups. This stage only nominates candidates; it cannot merge them.
Each candidate group is judged against its source evidence. Names merge only when the supplied records establish exact identity, not merely because the terms seem related. The build merged 193 groups and absorbed 209 duplicate entity records; uncertain cases remain separate.
Relationship names are combined only when they are interchangeable. Broader and narrower meanings stay distinct. Resolved entity names are then applied to every claim, duplicate edges are consolidated without erasing qualifier differences, and the searchable graph is produced.
Used for structured data extraction and graph resolution.
Used under human direction to consolidate duplicate entities, validate the extracted data, and correct errors.
The source states that high levels of calcium can slow healing.
The condition stays on the relationship. It does not become a separate entity called “high calcium,” and it is not discarded.
The graph is an automated interpretation of source material, not an independent scientific review. Extraction can miss context, choose the wrong direction, overstate a relationship or fail to recognize two names as the same thing. The source itself may also be incomplete, disputed or outdated. Use the rationale and source links to inspect important claims directly.