The invention discloses a
large model-based dialogue method in the field of
chemical industry, which comprises the following steps of: analyzing input data to identify equipment and molecular entities, and combining a pre-constructed equipment map and a
molecular model library to generate a topological embedding vector and chemical characteristics; comprehensive similarity is calculated based on the
semantic vector, the equipment correlation degree and the molecular similarity, and theme tags are dynamically inherited or given; analyzing an intention and matching an intention
label through a theme
label and a historical session, and triggering an agent response mechanism according to whether internal
database calling is involved or not: directly matching a pre-training agent with a non-
database scene to generate a response, extracting feature parameters from a
database scene through an
SQL template, fusing the feature parameters and returning data, and outputting a
natural language after agent reasoning. According to the method, equipment-molecule-process cross-
modal alignment is realized, the
response delay is reduced, and the professional intention recognition accuracy is improved.