Changes nemotron model

README.md

@@ -25,7 +25,7 @@ tags:
   - cosine-similarity
 models:
   - meta-llama/Llama-3.2-11B-Vision-Instruct
-  - nvidia/Nemotron-Mini-4B-Instruct
+  - nvidia/nemotron-3-nano-30b-a3b
   - sentence-transformers/all-MiniLM-L6-v2
 ---
 
@@ -64,7 +64,7 @@ flowchart TD
     A["📷 Photo from phone"] --> B["🤖 Vision AI\nLlama-3.2-11B-Vision\n(NVIDIA NIM / HF)"]
     B -->|"is it an animal?"| C{Animal detected?}
     C -->|"No"| ERR["❌ Error message\n'No dog or cat identified'"]
-    C -->|"Yes"| D["📝 Structured JSON\nNemotron Mini 4B Instruct structures\nspecies · breed · color · condition · marks"]
+    C -->|"Yes"| D["🧠 Nemotron 3 Nano (brain)\nreasons over the description →\nstructured JSON: species · breed · color · condition · marks"]
     D --> E["🔢 Semantic embedding\nall-MiniLM-L6-v2 · 384 dim\n(local, 22M params)"]
     E --> F{Cosine similarity\n≥ 0.80?}
     F -->|"Yes — known animal"| G["➕ New sighting\nadded to existing animal"]
@@ -106,14 +106,14 @@ The map uses color to signal urgency: **🟢 green** = dog · **🟠 orange** =
 
 | Component | Model / Library | Where it runs |
 |---|---|---|
-| Visual identification | **Llama-3.2-11B-Vision-Instruct** (11B) via NVIDIA NIM (or HF Serverless) | remote API |
-| Text reasoning / JSON | **Nemotron Mini 4B Instruct** via NVIDIA NIM | remote API |
+| Visual perception | **Llama-3.2-11B-Vision-Instruct** (11B) via NVIDIA NIM (or HF Serverless) | remote API |
+| Reasoning brain (structuring + animal stories) | **Nemotron 3 Nano** (`nvidia/nemotron-3-nano-30b-a3b`, MoE — 3B active) via NVIDIA NIM | remote API |
 | Semantic embedding | **all-MiniLM-L6-v2** (384 dim, 22M params) | CPU, local |
 | Animal matching | Cosine similarity (threshold 0.80) | CPU, local |
 | Database | SQLite | local / persistent |
 | Frontend | SPA via `gradio.Server` + Leaflet.js + Lucide Icons | browser |
 
-**Total parameters running locally: ~22M** (MiniLM). The vision and reasoning models run via serverless API — nothing heavy on the Space machine. **Every model used is well under the 32B hackathon cap:** Llama-3.2-11B-Vision (11B), Nemotron Mini 4B Instruct (4B) and MiniLM-L6-v2 (22M).
+**Total parameters running locally: ~22M** (MiniLM). Vision and the Nemotron brain run via serverless NVIDIA NIM — nothing heavy on the Space machine. **The hackathon cap is per model, not summed** ([per the FAQ](https://build-small-hackathon-field-guide.hf.space/#faq)), and every model is under 32B: Llama-3.2-11B-Vision (11B), Nemotron 3 Nano (30B total but only **3B active** per call, served serverless) and MiniLM-L6-v2 (22M).
 
 ---
 
@@ -136,8 +136,7 @@ This submission earns the following hackathon bonus quests:
 
 - 🎨 **Off-Brand** — Fully custom interface via `gradio.Server`, no default Gradio components visible. SPA frontend with Leaflet.js for the map, Lucide Icons, and custom design.
 - 📡 **Sharing is Caring** — Full traces of every sighting (photo → AI analysis → embedding → matching result) are automatically published as a dataset on the Hub via `/admin/push-traces`.
-- 🤏 **Tiny Titan** — The only on-device model is `all-MiniLM-L6-v2` (22M params, far under 4B): all animal matching runs locally on a tiny embedding model.
-- 🟩 **NVIDIA / Nemotron** — Structured identification is normalized by `Nemotron-Mini-4B-Instruct` via NVIDIA NIM — small (4B), well under the 32B cap.
+- 🟩 **NVIDIA / Nemotron** — **Nemotron 3 Nano** (`nvidia/nemotron-3-nano-30b-a3b`, MoE with 3B active) is the reasoning brain of the app: it structures every animal identification into the canonical record that drives matching, **and** writes each animal's story shown on its profile. The vision model only perceives the photo — Nemotron does the thinking.
 
 ---
 
@@ -148,7 +147,7 @@ git clone https://huggingface.co/spaces/build-small-hackathon/viralata-mapper
 cd viralata-mapper
 pip install -r requirements.txt
 
-# Recommended: NVIDIA NIM powers both vision (Llama-3.2-11B-Vision) and Nemotron Mini 4B
+# Recommended: NVIDIA NIM powers both vision (Llama-3.2-11B-Vision) and the Nemotron 3 Nano brain
 NVIDIA_API_KEY=nvapi_... python app.py
 
 # Alternative: vision via HuggingFace Serverless
@@ -170,7 +169,7 @@ On first run the app seeds the database with demo animals around Brasília so th
 
 | Secret | Description |
 |---|---|
-| `NVIDIA_API_KEY` | NVIDIA NIM key — powers **both** vision (Llama-3.2-11B-Vision) and Nemotron Mini 4B normalization. A single key is enough. |
+| `NVIDIA_API_KEY` | NVIDIA NIM key — powers **both** vision (Llama-3.2-11B-Vision) and the Nemotron 3 Nano brain (structuring + stories). A single key is enough. |
 | `HF_TOKEN` | Optional — vision via HF Serverless if you don't use NVIDIA NIM |
 | `MATCH_THRESHOLD` | Optional. Similarity threshold. Default: `0.80` |
 | `HF_DATASET_ID` | Optional. Hub dataset for trace publishing (e.g. `org/dataset-name`) |
@@ -223,7 +222,7 @@ Set up a **Persistent Storage Bucket** in the Space and mount it at `/data/` so
 ## Credits
 
 - **[Meta](https://ai.meta.com/llama/)** — Llama 3.2 11B Vision Instruct (Llama 3.2 Community License).
-- **[NVIDIA](https://build.nvidia.com/)** — Nemotron Mini 4B Instruct via NVIDIA NIM.
+- **[NVIDIA](https://build.nvidia.com/)** — Nemotron 3 Nano (`nvidia/nemotron-3-nano-30b-a3b`) via NVIDIA NIM.
 - **[Hugging Face](https://huggingface.co/)** — `sentence-transformers/all-MiniLM-L6-v2`, Serverless Inference, Space hosting.
 - **[Leaflet.js](https://leafletjs.com/)** — interactive map.
 - **[Gradio](https://gradio.app/)** — `gradio.Server` for the custom frontend.
@@ -233,7 +232,7 @@ Set up a **Persistent Storage Bucket** in the Space and mount it at `/data/` so
 
 ## License
 
-MIT for the application code. Models follow their own licenses — see the Meta (Llama 3.2) and NVIDIA (Nemotron Mini 4B Instruct) model cards for full terms.
+MIT for the application code. Models follow their own licenses — see the Meta (Llama 3.2) and NVIDIA (Nemotron 3 Nano / NVIDIA Open Model License) model cards for full terms.
 
 ---
 

app.py

@@ -46,6 +46,10 @@ def _photo_url(photo_path: str) -> str:
 # In-memory session store for analyze → confirm two-step flow
 _pending: dict[str, dict] = {}
 
+# Cache da historia gerada pelo Nemotron, por (animal_id, n_eventos).
+# Re-gera so quando ha avistamento/ajuda novo; evita chamada de API a cada visita.
+_story_cache: dict[tuple[int, int], str] = {}
+
 app = Server()
 
 # Serve photos as static files at /photos/...
@@ -97,7 +101,36 @@ async def get_animal(animal_id: int):
         s["photo_url"] = _photo_url(s.get("photo_path") or "")
     for h in detail.get("help_events", []):
         h["photo_url"] = _photo_url(h.get("photo_path") or "")
-    detail.get("animal", {}).pop("embedding", None)
+    animal = detail.get("animal", {})
+    animal.pop("embedding", None)
+
+    # Historia do animal redigida pelo Nemotron 3 Nano (cerebro do app).
+    # Cacheada por (id, n_eventos) e degrada para "" sem quebrar o perfil.
+    try:
+        sightings   = detail.get("sightings", [])
+        help_events = detail.get("help_events", [])
+        n_events    = len(sightings) + len(help_events)
+        cache_key   = (animal_id, n_events)
+        if cache_key in _story_cache:
+            detail["story"] = _story_cache[cache_key]
+        else:
+            desc_obj = json.loads(animal.get("description") or "{}")
+            location_hint = next(
+                (f"Lat {s['lat']:.3f}, Lng {s['lng']:.3f}"
+                 for s in sightings if s.get("lat") and s.get("lng")),
+                "",
+            )
+            story = ai.generate_story(
+                desc_obj,
+                sighting_count=len(sightings),
+                help_count=len(help_events),
+                location_hint=location_hint,
+            )
+            _story_cache[cache_key] = story
+            detail["story"] = story
+    except Exception:
+        detail["story"] = ""
+
     return JSONResponse(content=detail)
 
 

core/ai.py

@@ -1,17 +1,22 @@
 """
-ai.py — Visao + estruturacao de texto, ambas via NVIDIA NIM (ou HF como alternativa).
+ai.py — Visao + raciocinio/estruturacao de texto, ambas via NVIDIA NIM (ou HF como alternativa).
 
-Arquitetura (todos os modelos < 32B):
-  - Visao  : meta/llama-3.2-11b-vision-instruct   (11B, NVIDIA NIM)  -> NVIDIA_API_KEY
-             (ou meta-llama/Llama-3.2-11B-Vision-Instruct via HF, se so houver HF_TOKEN)
-  - Texto  : nvidia/nemotron-mini-4b-instruct      (4B,  NVIDIA NIM)  -> NVIDIA_API_KEY
-  - Embed  : sentence-transformers/all-MiniLM-L6-v2 (22M, local/CPU)
+Arquitetura (cada modelo individualmente < 32B — teto e por modelo, nao somado):
+  - Visao   : meta/llama-3.2-11b-vision-instruct    (11B, NVIDIA NIM)  -> NVIDIA_API_KEY
+              (ou meta-llama/Llama-3.2-11B-Vision-Instruct via HF, se so houver HF_TOKEN)
+  - Cerebro : nvidia/nemotron-3-nano-30b-a3b         (MoE, 3B ATIVOS,  NVIDIA NIM) -> NVIDIA_API_KEY
+  - Embed   : sentence-transformers/all-MiniLM-L6-v2 (22M, local/CPU)
+
+O Nemotron 3 Nano e o nucleo do app: alem de extrair o JSON estrito a partir da
+descricao livre da visao (especie, raca da lista permitida, cor, condicao, frase de
+matching), ele tambem REDIGE a "historia" de cada animal exibida no perfil
+(generate_story). E um modelo de reasoning MoE (30B totais / 3B ativos por chamada,
+256k de contexto) servido na API hospedada da NVIDIA — so 3B ativos o tornam rapido
+e barato, e o total de 30B fica abaixo do teto de 32B por modelo do hackathon.
 
 Fluxo: o Llama-3.2-11B-Vision enxerga a foto e descreve o animal (texto livre —
-modelos de visao raramente cospem JSON limpo). Em seguida o Nemotron Mini 4B
-Instruct (texto puro) LE essa descricao e EXTRAI o JSON estrito (especie, raca da
-lista permitida, cor, condicao, frase de matching). Cada modelo no que e bom:
-Llama ve, Nemotron estrutura.
+modelos de visao raramente cospem JSON limpo). Em seguida o Nemotron 3 Nano LE essa
+descricao, raciocina e estrutura. Cada modelo no que e bom: Llama ve, Nemotron pensa.
 
 Basta a NVIDIA_API_KEY: ela serve tanto para a visao quanto para o Nemotron.
 Sem NVIDIA_API_KEY mas com HF_TOKEN, a visao usa o HF Serverless e a estruturacao
@@ -21,7 +26,8 @@ Variaveis de ambiente:
   NVIDIA_API_KEY      — chave NVIDIA NIM (free tier em build.nvidia.com)
   HF_TOKEN            — token HuggingFace (alternativa de visao via HF Serverless)
   NVIDIA_VISION_MODEL — override do modelo de visao NIM (default meta/llama-3.2-11b-vision-instruct)
-  NVIDIA_MODEL        — override do modelo de texto NIM (default nvidia/nemotron-mini-4b-instruct)
+  NVIDIA_MODEL        — override do cerebro de texto NIM (default nvidia/nemotron-3-nano-30b-a3b;
+                        aponte para um Nemotron 3 Nano 4B self-hosted se rodar local)
   HF_VISION_MODEL     — override do modelo de visao HF (default meta-llama/Llama-3.2-11B-Vision-Instruct)
 """
 import base64
@@ -37,7 +43,7 @@ log = logging.getLogger(__name__)
 
 _HF_MODEL         = "meta-llama/Llama-3.2-11B-Vision-Instruct"  # id HF Serverless
 _NIM_VISION_MODEL = "meta/llama-3.2-11b-vision-instruct"        # id NVIDIA NIM (11B, free)
-_NIM_TEXT_MODEL   = "nvidia/nemotron-mini-4b-instruct"          # id NVIDIA NIM (4B, free)
+_NIM_TEXT_MODEL   = "nvidia/nemotron-3-nano-30b-a3b"            # id NVIDIA NIM (Nemotron 3 Nano MoE, 3B ativos)
 
 _DOG_BREEDS = (
     "SRD, Labrador Retriever, Golden Retriever, German Shepherd, Bulldog, French Bulldog, "
@@ -210,11 +216,16 @@ class AnimalAI:
             return self._fallback()
 
     def _structure_with_nemotron(self, raw_text: str) -> dict:
-        """Extrai JSON estrito da descricao livre da visao usando Nemotron Mini 4B Instruct."""
+        """Extrai JSON estrito da descricao livre da visao usando Nemotron 3 Nano."""
         try:
             resp = self.nim_client.chat.completions.create(
                 model=self.nim_text_model,
-                messages=[{"role": "user", "content": EXTRACT_PROMPT + raw_text}],
+                # "detailed thinking off": convencao do Nemotron 3 p/ suprimir traços de
+                # reasoning e devolver so a resposta — mantem o JSON limpo e a chamada barata.
+                messages=[
+                    {"role": "system", "content": "detailed thinking off"},
+                    {"role": "user", "content": EXTRACT_PROMPT + raw_text},
+                ],
                 max_tokens=500,
                 temperature=0.2,
                 top_p=0.9,
@@ -223,7 +234,7 @@ class AnimalAI:
             out = (msg.content or "") or (getattr(msg, "reasoning_content", "") or "")
             parsed = self._extract_json(out)
             if parsed is not None:
-                log.info("Nemotron Mini 4B estruturou o JSON.")
+                log.info("Nemotron 3 Nano estruturou o JSON.")
                 parsed["_ai_success"] = True
                 return parsed
             log.warning("Nemotron nao retornou JSON — tentando parse direto da visao")
@@ -237,6 +248,55 @@ class AnimalAI:
             return direct
         return self._fallback()
 
+    def generate_story(self, animal: dict, sighting_count: int,
+                       help_count: int = 0, location_hint: str = "") -> str:
+        """Redige uma breve historia do animal (1-2 frases) com o Nemotron 3 Nano.
+
+        E o cerebro do app dando voz ao perfil: a partir dos dados estruturados e do
+        historico, escreve uma narrativa curta em portugues para a pagina do animal.
+        Retorna "" se o Nemotron nao estiver disponivel ou falhar (o frontend so
+        mostra a historia quando ela existe — degrada sem quebrar).
+        """
+        if self.nim_client is None:
+            return ""
+
+        desc = animal if isinstance(animal, dict) else {}
+        facts = {
+            "species":       desc.get("species", ""),
+            "breed":         desc.get("breed_estimate", ""),
+            "color":         desc.get("primary_color", ""),
+            "condition":     desc.get("condition", ""),
+            "marks":         desc.get("distinctive_marks", []),
+            "sightings":     sighting_count,
+            "help_events":   help_count,
+            "location_hint": location_hint,
+        }
+        prompt = (
+            "You write a SHORT story (1-2 sentences, max 40 words) in BRAZILIAN PORTUGUESE "
+            "for the public profile of a stray animal being tracked by a community map. "
+            "Warm, factual, no invented details. Output ONLY the sentence(s), no preamble.\n"
+            "FACTS (JSON):\n" + json.dumps(facts, ensure_ascii=False)
+        )
+        try:
+            resp = self.nim_client.chat.completions.create(
+                model=self.nim_text_model,
+                messages=[
+                    {"role": "system", "content": "detailed thinking off"},
+                    {"role": "user", "content": prompt},
+                ],
+                max_tokens=160,
+                temperature=0.6,
+                top_p=0.9,
+            )
+            msg = resp.choices[0].message
+            text = (msg.content or "").strip()
+            # Modelo de reasoning pode emitir bloco <think>...</think> — remove.
+            text = re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL).strip()
+            return text
+        except Exception as e:
+            log.warning("Nemotron generate_story falhou: %s", e)
+            return ""
+
     def get_embedding(self, description: dict) -> list:
         """Embedding da descricao. Aleatorio se IA falhou (evita falsos matches)."""
         if not description.get("_ai_success", True):

index.html

@@ -581,7 +581,7 @@
     <h3>Tech stack</h3>
     <div class="stack-list">
       <div class="stack-row"><span class="stack-label">Vision AI</span><span class="stack-value">Llama 3.2 11B Vision Instruct</span></div>
-      <div class="stack-row"><span class="stack-label">Embeddings</span><span class="stack-value"> NVIDIA Nemotron Mini 4B Instruct</span></div>
+      <div class="stack-row"><span class="stack-label">Reasoning brain</span><span class="stack-value">NVIDIA Nemotron 3 Nano</span></div>
       <div class="stack-row"><span class="stack-label">Matching</span><span class="stack-value">all-MiniLM-L6-v2</span></div>
       <div class="stack-row"><span class="stack-label">Map</span><span class="stack-value">Leaflet.js + OSM</span></div>
       <div class="stack-row"><span class="stack-label">Frontend</span><span class="stack-value">gr.Server SPA</span></div>

setup-git.sh

@@ -37,7 +37,7 @@ git add .
 git commit -m "feat: initial Animal Visto app
 
 - Gradio UI mobile-first com mapa Leaflet
-- Integração NVIDIA NIM (Nemotron Nano VL)
+- Integração NVIDIA NIM (Nemotron 3 Nano + Llama 3.2 Vision)
 - Matching por cosine similarity (sentence-transformers)
 - SQLite persistente via HF Storage Bucket"
 

static/app.js

@@ -323,6 +323,25 @@
       const statusText = daysSince === 0 ? 'Seen today' : daysSince === 1 ? 'Seen yesterday' : `Seen ${daysSince} days ago`;
       document.getElementById('profile-status-text').textContent = statusText;
 
+      // História redigida pelo Nemotron 3 Nano (cérebro do app). Aparece só quando existe.
+      const story = (data.story || '').trim();
+      let storyEl = document.getElementById('profile-story');
+      const statusEl = document.getElementById('profile-status-text');
+      if (story) {
+        if (!storyEl) {
+          storyEl = document.createElement('p');
+          storyEl.id = 'profile-story';
+          storyEl.style.cssText = 'margin:.5rem 0 0;font-style:italic;opacity:.85;line-height:1.4;';
+          if (statusEl && statusEl.parentNode) {
+            statusEl.parentNode.insertBefore(storyEl, statusEl.nextSibling);
+          }
+        }
+        storyEl.textContent = story;
+        storyEl.style.display = '';
+      } else if (storyEl) {
+        storyEl.style.display = 'none';
+      }
+
       // Identification
       document.getElementById('prof-species').innerHTML = `${svgIcon(isDog?'dog':'cat',16)} ${isDog ? 'Dog' : 'Cat'}`;
       document.getElementById('prof-breed').textContent   = breed;