File size: 8,948 Bytes
089d665
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
"""Multimodal clinical extraction — single image → structured entities.

Lives in the Python backend (rather than the Next.js serverless function)
because:
  1. We can re-use `gemeo.extractor._kg_normalize_phenotypes` to map
     free-text labels to HP:xxxxxxx IDs via the raras-app KG
     phenotype_search fulltext (PT-BR coverage). The Vercel function
     can't reach Neo4j directly without extra latency.
  2. AI logic stays centralized — same auth/audit/redaction pipeline.
  3. Render gives fixed-cost compute; Vercel function execution costs
     scale per-call with vision payloads (~600KB each).

Provider chain — same model as the Next.js path (Groq Llama 4 Scout)
plus an optional Gemini fallback if the GEMINI_API_KEY is healthy.
Both share the same response shape so the front-end is agnostic to
which provider answered.

Public API:
  await extract_image(image_bytes, mime, source_url=None) -> dict
"""
from __future__ import annotations

import json
import logging
import os
from base64 import b64encode
from typing import Any

import httpx

logger = logging.getLogger("gemeo.multimodal_extract")


SYSTEM_PROMPT = """Você é um extrator clínico para um sistema de doenças raras.
Recebe uma imagem (screenshot de prontuário eletrônico, PDF, laudo, planilha)
e extrai entidades estruturadas em JSON.

Regras:
- Idioma fonte é provavelmente PT-BR — preserve termos clínicos no
  original em "label" e adicione tradução EN só se óbvia.
- Mapeie fenótipos a HPO IDs (HP:xxxxxxx) quando confiante. Se não
  conseguir mapear com >70% de confiança, omita o "id".
- Mapeie diagnósticos a ICD-10-BR e/ou ORPHA quando confiante.
- "confidence" 0..1 conforme certeza da extração (visibilidade do
  texto, ambiguidade clínica, qualidade da imagem).
- "evidence" copia a frase fonte que sustenta o achado — máximo 80 chars.
- Se a imagem não contém dado clínico, retorne arrays vazios e
  free_text descrevendo o que viu.
- NÃO INVENTE dados. Se um campo não está visível, omita-o.

Retorne APENAS um objeto JSON com este schema (campos vazios viram
array/objeto vazio, jamais null):
{
  "hpo": [{"id": "HP:xxxxxxx", "label": "...", "confidence": 0..1, "evidence": "..."}],
  "medications": [{"name": "...", "dose": "...", "route": "...", "confidence": 0..1}],
  "diagnoses": [{"name": "...", "icd10": "...", "orpha": "...", "confidence": 0..1}],
  "labs": [{"name": "...", "value": "...", "unit": "...", "date": "...", "confidence": 0..1}],
  "patient": {"age": "...", "sex": "...", "weight": "..."},
  "free_text": "...",
  "language": "pt-BR"
}"""


GROQ_URL = "https://api.groq.com/openai/v1/chat/completions"
GROQ_MODEL_DEFAULT = "meta-llama/llama-4-scout-17b-16e-instruct"

GEMINI_URL_TPL = (
    "https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={key}"
)
GEMINI_MODEL_DEFAULT = "gemini-2.5-flash"


def _normalize_entities(raw: dict[str, Any]) -> dict[str, Any]:
    """Cope with the model emitting Portuguese keys when the schema
    hint loses fidelity. Map alternates back to canonical keys + ensure
    arrays exist so the front-end never crashes on a missing key."""
    def pick_list(*keys: str) -> list:
        for k in keys:
            v = raw.get(k)
            if isinstance(v, list):
                return v
        return []

    return {
        "hpo": pick_list("hpo", "achados", "phenotypes", "fenotipos"),
        "medications": pick_list("medications", "medicamentos", "meds", "drugs"),
        "diagnoses": pick_list("diagnoses", "diagnosticos", "hipoteses", "differentials"),
        "labs": pick_list("labs", "exames", "laboratory", "tests"),
        "patient": raw.get("patient") or raw.get("paciente") or {},
        "free_text": (
            raw.get("free_text")
            if isinstance(raw.get("free_text"), str)
            else (raw.get("observacao") if isinstance(raw.get("observacao"), str) else "")
        ),
        "language": raw.get("language") if isinstance(raw.get("language"), str) else "pt-BR",
    }


def _parse_json_safely(text: str) -> dict[str, Any]:
    cleaned = text.strip()
    # Strip optional markdown code fences.
    if cleaned.startswith("```"):
        cleaned = cleaned.split("```", 2)[1] if "```" in cleaned[3:] else cleaned[3:]
        if cleaned.startswith("json"):
            cleaned = cleaned[4:]
        cleaned = cleaned.strip().rstrip("`").strip()
    return json.loads(cleaned)


async def _call_groq(image_b64: str, mime: str, user_prompt: str) -> dict[str, Any]:
    """Groq Llama 4 Scout vision — primary provider. ~600ms p50."""
    api_key = os.getenv("GROQ_API_KEY")
    if not api_key:
        raise RuntimeError("GROQ_API_KEY not set")
    model = os.getenv("GROQ_EXTRACT_MODEL", GROQ_MODEL_DEFAULT)
    data_url = f"data:{mime};base64,{image_b64}"
    payload = {
        "model": model,
        "messages": [
            {"role": "system", "content": SYSTEM_PROMPT},
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": user_prompt},
                    {"type": "image_url", "image_url": {"url": data_url}},
                ],
            },
        ],
        "response_format": {"type": "json_object"},
        "temperature": 0.1,
        "max_tokens": 4096,
    }
    async with httpx.AsyncClient(timeout=60.0) as client:
        r = await client.post(
            GROQ_URL,
            headers={
                "Authorization": f"Bearer {api_key}",
                "Content-Type": "application/json",
            },
            json=payload,
        )
        r.raise_for_status()
        j = r.json()

    text = j.get("choices", [{}])[0].get("message", {}).get("content", "")
    entities = _normalize_entities(_parse_json_safely(text))
    usage = j.get("usage") or {}
    return {
        "entities": entities,
        "model": f"groq:{model}",
        "tokens": {
            "input": usage.get("prompt_tokens"),
            "output": usage.get("completion_tokens"),
        },
    }


async def _call_gemini(image_b64: str, mime: str, user_prompt: str) -> dict[str, Any]:
    """Gemini 2.5 Flash vision — fallback. Uses responseSchema for
    strict JSON. Native PT-BR; handles tabular reports a touch better
    than Llama 4 Scout but ~3× slower."""
    api_key = os.getenv("GEMINI_API_KEY")
    if not api_key:
        raise RuntimeError("GEMINI_API_KEY not set")
    model = os.getenv("GEMINI_EXTRACT_MODEL", GEMINI_MODEL_DEFAULT)
    url = GEMINI_URL_TPL.format(model=model, key=api_key)
    payload = {
        "systemInstruction": {"parts": [{"text": SYSTEM_PROMPT}]},
        "contents": [
            {
                "role": "user",
                "parts": [
                    {"text": user_prompt},
                    {"inlineData": {"mimeType": mime, "data": image_b64}},
                ],
            }
        ],
        "generationConfig": {
            "responseMimeType": "application/json",
            "temperature": 0.1,
            "maxOutputTokens": 4096,
        },
    }
    async with httpx.AsyncClient(timeout=60.0) as client:
        r = await client.post(url, json=payload)
        r.raise_for_status()
        j = r.json()
    text = j.get("candidates", [{}])[0].get("content", {}).get("parts", [{}])[0].get("text", "")
    entities = _normalize_entities(_parse_json_safely(text))
    usage = j.get("usageMetadata") or {}
    return {
        "entities": entities,
        "model": f"gemini:{model}",
        "tokens": {
            "input": usage.get("promptTokenCount"),
            "output": usage.get("candidatesTokenCount"),
        },
    }


async def extract_image(
    image_bytes: bytes,
    mime: str = "image/png",
    source_url: str | None = None,
) -> dict[str, Any]:
    """Run a screenshot / clinical image through a multimodal model and
    return canonical structured entities.

    Provider order: Groq Llama 4 Scout → Gemini Flash. Returns
    `{entities, model, tokens, fallback?, elapsed_ms}` so the caller can
    log which provider answered.
    """
    import time

    started = time.time()
    image_b64 = b64encode(image_bytes).decode("ascii")
    user_prompt = (
        f"Captura de tela vinda de: {source_url}. Extraia as entidades clínicas."
        if source_url
        else "Extraia as entidades clínicas desta imagem."
    )

    errors: list[str] = []
    for fn, label in [(_call_groq, "groq"), (_call_gemini, "gemini")]:
        try:
            result = await fn(image_b64, mime, user_prompt)
            result["elapsed_ms"] = int((time.time() - started) * 1000)
            if label != "groq":
                result["fallback"] = True
            return result
        except Exception as e:
            msg = str(e)[:240]
            logger.warning("[multimodal_extract] %s failed: %s", label, msg)
            errors.append(f"{label}: {msg}")

    raise RuntimeError(f"all providers failed — {' | '.join(errors)}")