Enable spatio-temporal parameters to be passed to API requests

[simeonwetzel]

In addition to text inputs, the API-endpoint for conversational chatbot should also be able to consume temporal or spatial constraints via the API requests.

[simeonwetzel]

Implemenation Details:

Temporal and spatial filters can now be passed to the chatbot API-endpoint.
Next to the text input (query) parameter it now allows an optional spatio_temporal_context param like:

{
    "query": "Can you show me building data for Dresden? I need hospitals",
    "spatio_temporal_context": {"temporal": "2020-2021","extent": [13.5793237, 51.1777202, 14.9660626, 50.974937]}
}

Usage:

curl -X POST http://localhost:8000/data \
-H "Content-Type: application/json" \
-d '{
    "query": "Can you show me building data for Dresden? I need hospitals",
    "spatio_temporal_context": {
        "temporal": "2020-2021",
        "extent": [13.5793237, 51.1777202, 14.9660626, 50.974937]
    }
}'

Endpoint implemenation details:

innovation-prize/search-app/server/app/server.py

Lines 150 to 167 in c1a3b0a

	class Query(BaseModel):
	query: str
	spatio_temporal_context: Optional[Dict[str, Any]] = None
	@app.post("/data")
	async def call_graph(query_data: Query, session_id: UUID = Depends(cookie)):
	if graph is not None:
	print(f"-#-#--Running graph---- Using session_id: {str(session_id)}")
	inputs = {"messages": [HumanMessage(content=query_data.query)]}
	if query_data.spatio_temporal_context:
	inputs['spatio_temporal_context'] = query_data.spatio_temporal_context
	graph.graph.thread_id = str(session_id)
	response = await graph.ainvoke(inputs)
	else:
	raise HTTPException(status_code=400, detail="No session created")
	return response

Additional feature: Automated derivation of spatial constraints from user input

The spatio_temporal_context parameter is optional. If users do not provide it, an LLM+tools is used to derive the spatial context automatically.
Example:
"query": "Can you show me building data for Dresden?"
Extraced spatial context:
{'name': 'Dresden', 'country': 'Deutschland', 'type': 'city', 'extent': [13.5793237, 51.1777202, 13.9660626, 50.974937]}

How it works:

1. We use an LLM for Named Entity Recognition (extract toponyms) and spatial scale ("Local", "City", "Regional", "National", "Continental", "Global"). Used prompt:

prompt="""You are an expert in geography and spatial data. 
    Your task is to extract from a query spatial entities such as city, country or region names.
    Also determine the spatial scale ("Local", "City", "Regional", "National", "Continental", "Global") from the given query.
"""

2. The geodoing service by Komoot is used to get a list of candidates with bounding boxes.
3. We use another LLM to get the most reasonable candidate of the komoot results. Used prompt:

prompt="""You are an expert in geography and spatial data. 
    Your task is to pick from the results list the best matching candidate according to the query.
    If the original query includes a country information, consider this in your selection.
    If also consider the type. E.g. if user asks for a 'river' also pick the corresponding result

    Also consider the scale: {scale}
    Query: {original_query}
    Results: {results}
    Output:"""