Search reaction evidence
Choose exact matching deliberately and read the selected source.
Match the intended object
| Question | Tool |
|---|---|
| Preparations of one molecule | molecule_reactions(smiles, role="product") |
| Uses of one molecule | molecule_reactions(smiles, role="reactant") |
| Complete reactant/product side plus conditions | search_reactions |
| The same full transformation | lookup_reaction |
| Structurally similar product sets | similar_products |
| Full source record | get_observation |
| Relationships between entities | graph |
products and reactants in search_reactions mean the complete dot-separated side. They are not substructure queries. An empty whole-side result does not rule out a molecule recorded with a coproduct; try a molecule query where appropriate.
Start with 3–5 results. Follow has_more and next_offset only when more evidence is useful. Keep different observations separate rather than averaging incompatible conditions.
Read conditions by stage
Temperature uses Celsius and pressure uses bar. Preserve raw units, intervals and pressure kind. extracted, predicted and unknown are different statuses. Never turn an unknown value into room temperature, zero or atmospheric pressure.
The source paragraph can distinguish reaction, neutralization, evaporation, drying and purification. A vacuum used for drying is not the reaction pressure. An acid added after conversion is not automatically its catalyst. roles_resolved:false means the agent list is not a resolved list of catalysts or solvents.
Follow provenance
The actual opaque observation_id identifies a record. evidence.uid, patent number and record_id locate the source, but cannot replace the observation ID in a lookup call. record_digest guards reviewed overrides against source changes; dataset_revision describes the corpus snapshot.
Accepted corrections appear with correction_provenance, original conditions and review versions. Raw source evidence stays unchanged. Accepted community contributions have evidence.kind="community_contribution"; they participate in exact retrieval. Structural similarity remains limited to the indexed base corpus, with no invented community similarity score.
Tanimoto similarity is structural resemblance, not a reaction-success probability. A relationship graph is not a validated multistep synthesis plan.