GeoAI Outdoor Assistant — intent parser and slot chooser

This repository is the model that runs inside GeoAI Outdoor Assistant, a GeoAI Hackathon @ CSRE, IIT Bombay submission. It answers free-form questions like "I have mild asthma, where's good for a morning walk near Powai in the next two days?" with a specific place, a specific time, and evidence for every claim.

The model we use is Qwen/Qwen3-1.7B (Apache-2.0), unchanged and quantized to Q4_K_M for CPU serving. This repository ships that quantized GGUF, plus the prompt contract, the JSON schemas, and the serving settings.

Provenance

This repository hosts the quantized build we actually serve, plus the configuration and documentation for how we use it. The unquantized weights stay in Qwen/Qwen3-1.7B; they are not copied here.

How the model is used

The model does two narrow jobs and nothing else. All facts come from live API calls, and the model never recalls weather or geography.

1. Intent parsing. Input is the user's question plus the current date. Output is a JSON object constrained to a schema:

{
  "activity": "running",
  "location": "JP Nagar 5th Phase",
  "day": "tomorrow",
  "part_of_day": "morning",
  "hour": null,
  "profile": ["asthma"],
  "out_of_scope": false
}

2. Slot choosing. Input is a table of already-scored options (place, time, score, temperature, air quality, rain chance). Output is a JSON object with two fields:

{
  "slot_id": "S1-102218",
  "headline": "Go to Anuroopa Park around 09:00 on Thu 22 Oct."
}

The model picks the slot and writes that one sentence. Everything else in the answer - the reasons with their evidence, the carry list, the watch-outs and the avoid windows - is composed in code from the same fetched data, so the model is never the source of a fact. That is deliberate: a small model that is asked for the reasons will eventually state a number that was never fetched.

The model is asked for JSON and the output is constrained by a grammar compiled from the schema, so the structure is always valid.

The exact schemas are published next to this card, so the contract is inspectable rather than described:

A separate rule-based verifier then re-checks every number, clock time, place name and evidence citation in the model's wording against the fetched data. If a check fails the model is asked once more; if it fails again the answer is produced by a deterministic template built from the same data. The model is never allowed to state a fact that was not fetched.

Intended use and limitations

• Intended use. Ranking and explaining shortlists of outdoor venues for an urban user in India, and extracting intent from free-form English questions about outdoor activities.
• Not for medical advice. Profile handling (asthma, senior, child, pregnancy) only tightens conservative thresholds. Every profile-adjusted answer carries a "not medical advice" note.
• Not a knowledge base. The model has no memory of weather, place names or AQI. It must be given the data.
• English only. The parser is tuned for English and light Hinglish. Other languages are out of scope.
• Small-model limits. The parser occasionally misreads an unusual phrasing. The production system compensates by deriving the calendar and the profile match deterministically in code, so the user's own wording always wins over the model's interpretation.

Loading

This repository is GGUF-only and loads with llama-cpp-python, the same runtime the service uses:

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="som-doe2007/geoai-outdoor-assistant",
    filename="Qwen3-1.7B-Q4_K_M.gguf",
    n_ctx=4096,
    n_threads=4,
    verbose=False,
)

Llama.from_pretrained(...) downloads the GGUF from this repository and loads it with llama.cpp. The unquantized upstream model, if you want to inspect or re-quantize it, loads with the standard transformers API:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-1.7B")

Qwen3 emits a reasoning block by default. We disable it by pre-seeding an empty <think>\n\n</think> block in the assistant turn, which is the officially supported way to switch thinking off, and which keeps the answer short enough for CPU inference.

This is not cosmetic. With thinking enabled the JSON grammar is not always enforced (llama.cpp issue 20345), and the answer would not be parseable. The guard is covered by a test in the application repository.

The GGUF in this repository

Qwen3-1.7B-Q4_K_M.gguf is a third-party quantisation of Qwen/Qwen3-1.7B by the unsloth organisation, Apache-2.0. It is byte-identical (SHA-256 b139949c5bd74937ad8ed8c8cf3d9ffb1e99c866c823204dc42c0d91fa181897) to the file the service loads, and it is the only weight file in this repository.

Chat template note

The repository uses the Qwen3 chat template. The template accepts enable_thinking; when it is false the template pre-seeds the empty think block itself, so the same behaviour is obtained either way.

Data sources

The model itself is trained on nothing from this project. The application that uses it fetches its data live at query time from free and open sources: OpenStreetMap (Overpass API, Nominatim), Open-Meteo (weather and air quality via CAMS ENSEMBLE), WAQI, OSRM and Open-Meteo/Astral for daylight.

Disclosure

• The only language model used at request time is this one, run locally. No GPT-4-class, Claude or Gemini model is called anywhere in the running assistant.

© OpenStreetMap contributors for the map data used by the application that serves this model.