AinoAI_v1

Aino AI

Aino AI is a LLM designed by Aino, a company specializing in smart GIS analysis. The model empowers both technical and non-technical users to perform comprehensive geospatial analyses with a focus on geometry processing.

Model Overview

Aino AI is a first generation model that specifies on a geospatial analysis with PostGIS. The model is a version of finetuned Mistral 7B for converting user prompts to SQL including spatial functions like ST_Buffer, ST_Distance, ST_Envelope and others.

Applications

Since AinoAI is used for complex tasks, it has strict input format:

{
'prompt': 'find all /restaurants in 200 m from metro stations from / public_transport_stations', 
	'tables':[
		'restaurants':[
			{'id': [..., ...]},
			{'name': [..., ...]}, 
			{'categories': [..., ...]}, 
			{'address': [..., ...]}, 
			{'geometry': [..., ...]}],
		'public_transport_stations':[
			{'id': [..., ...]}, 
			{'transport_type': [..., ...]}, 
			{'address': [..., ...]}, 
			{'geometry': [..., ...]}],
			
	]
}

Prompt is user part where one can specify needed operation and tables is autogenerated part created by application itself bacing on links to datasets involved in request. Table section contains list of objects named by dataset and list of columns inside dataset with examples. The model is trained in a way to recognize data formats of columns and assume needed parameters of filtration, so even for huge datasets we pass up to 10 examples which is enough.

The response of a model is a SQL response with varying placeholders that must be replace if the are presented. Those are: • CRS – must be replaced for convertation betwen ESPG:4326 to local metric system and is used in prompts like – calculate distance in meters, area of objects in datasets and others, where processing must be in meters • {mock_geometry} – is a placeholder for user geometry, and is used in prompts like “find green spaces in /user_polygon”, since user geometry in Aino is stored in WKT format which has no influence of SQL, we don`t process geometry itself for saving time for SQL generation and expanding input data.

The response after postprocessing is running in a background and creates dataset as a response, that is visualized in application.

Although the model is trained to understand short prompts, for accurate results we suggest users to formulate them in a proper way and specify any relevant information.

Prompting

Model is focused on work with geospatial data, and as you may know OSM is biggest library of such data. Here are some tips on how to work with such data. For example names can be presented in columns: name, name_en, name_esp, name_fr, so if user needs to select records by a name, it is better to specify which language it is.

Same for requesting filtered data, in case of usage prompt like “List all café in /Eixample_Barcelona” user will get list of names and changing “List all” to “Find all” will drive to getting filtered dataset with point locations.

Here you can see real examples of combinations of input and output:

Training Data

AinoAI was trained on over 10,000 self-created sequences of user requests and table schemas. This robust dataset ensures the model’s ability to generate precise SQL queries tailored to user-provided schemas. Benchmarking

Test dataset consists of examples to evaluate SQL creation for filtering by parameters, geometry processing, join operators for processing list of tables at onetime and math operations – ratio of elements and geometries, percentage. Since other LLM in list are perfect for chat-like interaction, geospatial analysis is relativelly narrow branch of knowledges and for precise results model need to be trained in a proper way with focus on speciffic.

Access

AinoAI is exclusively available through the Aino application, accessible via subscription. For more information, visit https://aino.world/.

Future Development

Aino is committed to enhancing AinoAI’s capabilities. Upcoming updates will focus on:

• Improved performance
• Support for staged logic queries, enabling more complex and multi-step geospatial analyses