-
⚔️
-
WHETSTONE
-
Loading training environment…
-
-
+
⚔️
+
WHETSTONE
+
Legal Training Dojo
+
-
From 5d7c70333e7d295a1e3708213c2d0b8097a1223b Mon Sep 17 00:00:00 2001
From: Vector897 <242271150+Vector897@users.noreply.github.com>
Date: Sat, 27 Jun 2026 13:54:14 +0100
Subject: [PATCH 2/2] GCP integration: Vertex AI Gemini + Cloud Run + BigQuery
(2026-06-27)
- Dockerfile + .dockerignore for Cloud Run deployment (europe-west2)
- llm.py: auto-route to Vertex AI Gemini when model starts with 'gemini'
and GCP_PROJECT is configured; OpenRouter remains fallback
- config.py: GCP_PROJECT, VERTEX_LOCATION; JUDGE_MODEL auto-defaults to
gemini-2.0-flash-001 when GCP_PROJECT is set
- data/bigquery.py: BigQuery corpus layer (whetstone.corpus table);
graceful fallback to SQLite when GCP not configured
- main.py: ensure_table() on startup; /api/health reports GCP status;
corpus search prefers BigQuery over SQLite
- requirements.txt: google-cloud-aiplatform>=1.60.0, google-cloud-bigquery>=3.11.0
- .gitignore: carve out backend/requirements.txt from *.txt rule
---
.dockerignore | 19 ++++++
.gitignore | 2 +
Dockerfile | 19 ++++++
backend/app/config.py | 25 ++++++--
backend/app/data/bigquery.py | 118 +++++++++++++++++++++++++++++++++++
backend/app/llm.py | 109 +++++++++++++++++++++++++-------
backend/app/main.py | 16 ++++-
backend/requirements.txt | 9 +++
8 files changed, 288 insertions(+), 29 deletions(-)
create mode 100644 .dockerignore
create mode 100644 Dockerfile
create mode 100644 backend/app/data/bigquery.py
create mode 100644 backend/requirements.txt
diff --git a/.dockerignore b/.dockerignore
new file mode 100644
index 0000000..36b0c73
--- /dev/null
+++ b/.dockerignore
@@ -0,0 +1,19 @@
+.venv/
+.env
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+data/*.db
+data/whetstone.db
+.git/
+*.md
+*.txt
+*.pdf
+.kaggle_ref/
+outputs/
+thesis/
+scripts/
+tests/
+0608/
+W2/
diff --git a/.gitignore b/.gitignore
index 05b843e..33392de 100644
--- a/.gitignore
+++ b/.gitignore
@@ -4,6 +4,8 @@
*.md
*.txt
*.pdf
+# exceptions — code files that happen to use .txt extension
+!backend/requirements.txt
# secrets
.env
diff --git a/Dockerfile b/Dockerfile
new file mode 100644
index 0000000..d1ad6c1
--- /dev/null
+++ b/Dockerfile
@@ -0,0 +1,19 @@
+FROM python:3.13-slim
+
+WORKDIR /app
+
+# Install Python dependencies
+COPY backend/requirements.txt requirements.txt
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Optional: Presidio NLP model (non-fatal if it fails)
+RUN python -m spacy download en_core_web_sm 2>/dev/null || true
+
+# Copy application source
+COPY . .
+
+# Cloud Run injects PORT; default 8080
+ENV PORT=8080
+EXPOSE 8080
+
+CMD exec uvicorn app.main:app --app-dir backend --host 0.0.0.0 --port ${PORT}
diff --git a/backend/app/config.py b/backend/app/config.py
index 9d7c464..fb06233 100644
--- a/backend/app/config.py
+++ b/backend/app/config.py
@@ -1,19 +1,34 @@
-"""集中读取 .env 配置（项目根 D:\\Whetstone\\.env）。"""
+"""Centralised config — reads .env at project root (D:\\Whetstone\\.env)."""
from pathlib import Path
import os
from dotenv import load_dotenv
-# backend/app/config.py -> 项目根 = parents[2]
ROOT = Path(__file__).resolve().parents[2]
load_dotenv(ROOT / ".env")
-OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY", "")
+# ── OpenRouter ────────────────────────────────────────────────────────────────
+OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY", "")
OPENROUTER_BASE_URL = os.getenv("OPENROUTER_BASE_URL", "https://openrouter.ai/api/v1")
+# ── Model routing ─────────────────────────────────────────────────────────────
+# ADVERSARY_MODEL — main opponent (Claude for best adversarial quality)
+# JUDGE_MODEL — scoring + concession-gate referee
+# set to "gemini-2.0-flash-001" to use Vertex AI (GCP prize track)
+# REDTEAM_MODEL — red/blue team debate
ADVERSARY_MODEL = os.getenv("ADVERSARY_MODEL", "anthropic/claude-3.7-sonnet")
-JUDGE_MODEL = os.getenv("JUDGE_MODEL", "nvidia/llama-3.1-nemotron-70b-instruct")
-REDTEAM_MODEL = os.getenv("REDTEAM_MODEL", "nvidia/llama-3.1-nemotron-70b-instruct")
+JUDGE_MODEL = os.getenv(
+ "JUDGE_MODEL",
+ # Default to Gemini via Vertex AI if GCP_PROJECT is configured, else Nemotron
+ "gemini-2.0-flash-001" if os.getenv("GCP_PROJECT") else "nvidia/llama-3.1-nemotron-70b-instruct",
+)
+REDTEAM_MODEL = os.getenv("REDTEAM_MODEL", "nvidia/llama-3.1-nemotron-70b-instruct")
+# ── GCP / Vertex AI ───────────────────────────────────────────────────────────
+GCP_PROJECT = os.getenv("GCP_PROJECT", "") # frictionless account project id
+VERTEX_LOCATION = os.getenv("VERTEX_LOCATION", "europe-west2")
+
+# ── EU law ────────────────────────────────────────────────────────────────────
CELLAR_SPARQL = os.getenv("CELLAR_SPARQL", "https://publications.europa.eu/webapi/rdf/sparql")
+# ── Paths ─────────────────────────────────────────────────────────────────────
SCENARIO_DIR = Path(__file__).resolve().parent / "scenarios"
diff --git a/backend/app/data/bigquery.py b/backend/app/data/bigquery.py
new file mode 100644
index 0000000..dc12779
--- /dev/null
+++ b/backend/app/data/bigquery.py
@@ -0,0 +1,118 @@
+"""BigQuery corpus layer — GCP prize track: 'A home for your data'.
+
+Stores SEC EDGAR / CUAD legal clause corpus in BigQuery alongside SQLite.
+Falls back gracefully to SQLite-only if BigQuery is unavailable or unconfigured.
+
+Usage:
+ from .data import bigquery as bq
+ bq.ensure_table() # called once on startup
+ bq.insert(rows) # list of dicts matching SCHEMA
+ bq.search("liability_cap") # returns list of clause dicts
+"""
+from __future__ import annotations
+import logging
+from typing import Any
+
+from .. import config
+
+log = logging.getLogger(__name__)
+
+DATASET = "whetstone"
+TABLE = "corpus"
+
+# ── Optional BigQuery import ──────────────────────────────────────────────────
+
+try:
+ from google.cloud import bigquery as _bq
+ _BQ_OK = True
+except ImportError:
+ _BQ_OK = False
+
+# ── Internal helpers ──────────────────────────────────────────────────────────
+
+_client_cache: Any = None
+
+
+def _client():
+ global _client_cache
+ if not _BQ_OK or not config.GCP_PROJECT:
+ return None
+ if _client_cache is None:
+ _client_cache = _bq.Client(project=config.GCP_PROJECT)
+ return _client_cache
+
+
+def _full_table() -> str:
+ return f"{config.GCP_PROJECT}.{DATASET}.{TABLE}"
+
+
+# ── Public API ────────────────────────────────────────────────────────────────
+
+def ensure_table() -> bool:
+ """Auto-create BigQuery dataset + corpus table. Returns True if BQ is active."""
+ c = _client()
+ if not c:
+ return False
+ try:
+ ds_ref = _bq.Dataset(f"{config.GCP_PROJECT}.{DATASET}")
+ ds_ref.location = config.VERTEX_LOCATION or "EU"
+ c.create_dataset(ds_ref, exists_ok=True)
+ schema = [
+ _bq.SchemaField("id", "STRING", mode="REQUIRED"),
+ _bq.SchemaField("source", "STRING"),
+ _bq.SchemaField("clause_type", "STRING"),
+ _bq.SchemaField("text", "STRING"),
+ _bq.SchemaField("metadata", "JSON"),
+ _bq.SchemaField("inserted_at", "TIMESTAMP",
+ default_value_expression="CURRENT_TIMESTAMP()"),
+ ]
+ tbl = _bq.Table(_full_table(), schema=schema)
+ c.create_table(tbl, exists_ok=True)
+ log.info("BigQuery table ready: %s", _full_table())
+ return True
+ except Exception as exc:
+ log.warning("BigQuery setup failed (will use SQLite fallback): %s", exc)
+ return False
+
+
+def insert(rows: list[dict]) -> bool:
+ """Stream-insert rows into BigQuery corpus table."""
+ c = _client()
+ if not c or not rows:
+ return False
+ try:
+ errors = c.insert_rows_json(_full_table(), rows)
+ if errors:
+ log.warning("BigQuery insert errors: %s", errors)
+ return len(errors) == 0
+ except Exception as exc:
+ log.warning("BigQuery insert failed: %s", exc)
+ return False
+
+
+def search(clause_type: str, limit: int = 10) -> list[dict]:
+ """Full-text search by clause_type in BigQuery. Returns [] on failure."""
+ c = _client()
+ if not c:
+ return []
+ try:
+ query = f"""
+ SELECT id, source, clause_type, text, metadata
+ FROM `{_full_table()}`
+ WHERE clause_type LIKE @clause_type
+ LIMIT @limit
+ """
+ job_config = _bq.QueryJobConfig(query_parameters=[
+ _bq.ScalarQueryParameter("clause_type", "STRING", f"%{clause_type}%"),
+ _bq.ScalarQueryParameter("limit", "INT64", limit),
+ ])
+ rows = c.query(query, job_config=job_config).result()
+ return [dict(r) for r in rows]
+ except Exception as exc:
+ log.warning("BigQuery search failed: %s", exc)
+ return []
+
+
+def active() -> bool:
+ """True if BigQuery is installed and GCP_PROJECT is configured."""
+ return _BQ_OK and bool(config.GCP_PROJECT)
diff --git a/backend/app/llm.py b/backend/app/llm.py
index 6be137a..2dee17c 100644
--- a/backend/app/llm.py
+++ b/backend/app/llm.py
@@ -1,4 +1,9 @@
-"""OpenRouter LLM 封装。统一 chat / JSON 解析 / CoT 开关。"""
+"""LLM client — OpenRouter (primary) + Vertex AI Gemini (GCP integration).
+
+Routing rule:
+ model starts with "gemini" AND GCP_PROJECT is set → Vertex AI
+ everything else → OpenRouter
+"""
from __future__ import annotations
import json
import re
@@ -6,15 +11,28 @@
from openai import OpenAI
from . import config
-_client: OpenAI | None = None
+# ── Optional Vertex AI import ─────────────────────────────────────────────────
+
+try:
+ import vertexai
+ from vertexai.generative_models import (
+ GenerativeModel, Content, Part, GenerationConfig,
+ )
+ _VERTEX_AVAILABLE = True
+except ImportError:
+ _VERTEX_AVAILABLE = False
+
+# ── OpenRouter client ─────────────────────────────────────────────────────────
+_or_client: OpenAI | None = None
-def client() -> OpenAI:
- global _client
- if _client is None:
+
+def _openrouter() -> OpenAI:
+ global _or_client
+ if _or_client is None:
if not config.OPENROUTER_API_KEY:
- raise RuntimeError("OPENROUTER_API_KEY 未配置（检查 .env）")
- _client = OpenAI(
+ raise RuntimeError("OPENROUTER_API_KEY not configured (.env)")
+ _or_client = OpenAI(
api_key=config.OPENROUTER_API_KEY,
base_url=config.OPENROUTER_BASE_URL,
default_headers={
@@ -22,28 +40,70 @@ def client() -> OpenAI:
"X-Title": "Whetstone",
},
)
- return _client
+ return _or_client
+
+
+# ── Vertex AI Gemini ──────────────────────────────────────────────────────────
+
+def _to_gemini_contents(messages: list[dict]) -> tuple[str, list]:
+ """Split system instruction from user/model turns."""
+ system_parts: list[str] = []
+ contents = []
+ for m in messages:
+ role = m.get("role", "")
+ text = m.get("content", "")
+ if role == "system":
+ system_parts.append(text)
+ elif role == "user":
+ contents.append(Content(role="user", parts=[Part.from_text(text)]))
+ elif role in ("assistant", "model"):
+ contents.append(Content(role="model", parts=[Part.from_text(text)]))
+ return "\n\n".join(system_parts), contents
+
+
+def _chat_vertex(messages: list[dict], model: str, temperature: float = 0.7,
+ max_tokens: int = 1200) -> str:
+ if not _VERTEX_AVAILABLE:
+ raise RuntimeError(
+ "google-cloud-aiplatform not installed — "
+ "pip install google-cloud-aiplatform>=1.60"
+ )
+ vertexai.init(project=config.GCP_PROJECT, location=config.VERTEX_LOCATION)
+ system_txt, contents = _to_gemini_contents(messages)
+ gem = GenerativeModel(
+ model,
+ system_instruction=system_txt or None,
+ )
+ resp = gem.generate_content(
+ contents,
+ generation_config=GenerationConfig(
+ temperature=temperature,
+ max_output_tokens=max_tokens,
+ ),
+ )
+ return resp.text.strip()
+
+# ── Unified interface ─────────────────────────────────────────────────────────
def chat(messages: list[dict], model: str | None = None, temperature: float = 0.7,
max_tokens: int = 1200) -> str:
- """返回文本。model 缺省走对手模型。"""
- resp = client().chat.completions.create(
- model=model or config.ADVERSARY_MODEL,
- messages=messages,
- temperature=temperature,
- max_tokens=max_tokens,
+ """Route to Vertex AI Gemini when model starts with 'gemini' and GCP is configured."""
+ m = model or config.ADVERSARY_MODEL
+ if m.startswith("gemini") and config.GCP_PROJECT:
+ return _chat_vertex(messages, model=m, temperature=temperature,
+ max_tokens=max_tokens)
+ resp = _openrouter().chat.completions.create(
+ model=m, messages=messages, temperature=temperature, max_tokens=max_tokens,
)
return (resp.choices[0].message.content or "").strip()
def _extract_json(text: str) -> Any:
- """从模型输出里抠出第一个 JSON 对象（容忍 ```json 包裹与前后噪声）。"""
text = text.strip()
fence = re.search(r"```(?:json)?\s*(\{.*?\}|\[.*?\])\s*```", text, re.DOTALL)
if fence:
text = fence.group(1)
- # 退化：找第一个 { 到最后一个 }
if not text.startswith("{") and not text.startswith("["):
s, e = text.find("{"), text.rfind("}")
if s != -1 and e != -1:
@@ -54,7 +114,7 @@ def _extract_json(text: str) -> Any:
def chat_json(messages: list[dict], model: str | None = None,
temperature: float = 0.3, max_tokens: int = 2000,
retries: int = 1) -> dict:
- """要求模型返回 JSON，解析为 dict。失败重试。"""
+ """Call chat() and parse result as JSON dict, retrying on parse failure."""
last_err = None
for _ in range(retries + 1):
raw = chat(messages, model=model, temperature=temperature, max_tokens=max_tokens)
@@ -62,7 +122,14 @@ def chat_json(messages: list[dict], model: str | None = None,
return _extract_json(raw)
except Exception as e: # noqa: BLE001
last_err = e
- messages = messages + [
- {"role": "user", "content": "上一条不是合法 JSON，请只输出一个合法 JSON 对象，不要解释。"}
- ]
- raise ValueError(f"JSON 解析失败: {last_err}")
+ messages = messages + [{
+ "role": "user",
+ "content": "Previous response was not valid JSON. "
+ "Reply with a single valid JSON object only, no explanation.",
+ }]
+ raise ValueError(f"JSON parse failed after retries: {last_err}")
+
+
+def vertex_available() -> bool:
+ """True if Vertex AI is installed and GCP_PROJECT is configured."""
+ return _VERTEX_AVAILABLE and bool(config.GCP_PROJECT)
diff --git a/backend/app/main.py b/backend/app/main.py
index 6a6a088..caa882f 100644
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -8,9 +8,9 @@
from fastapi.responses import FileResponse
from pydantic import BaseModel
-from . import config, state, db
+from . import config, state, db, llm
from .engines import adversary, scoring, redteam, desensitize
-from .data import cellar, edgar
+from .data import cellar, edgar, bigquery as bq
# ── App setup ─────────────────────────────────────────────────────────────────
@@ -24,6 +24,7 @@
@app.on_event("startup")
def startup():
db.init()
+ bq.ensure_table() # no-op if GCP_PROJECT not set
# ── Helpers ───────────────────────────────────────────────────────────────────
@@ -54,6 +55,13 @@ def health():
"ok": True,
"key_configured": bool(config.OPENROUTER_API_KEY),
"corpus_size": db.corpus_count(),
+ "gcp": {
+ "project": config.GCP_PROJECT or None,
+ "vertex_ai": llm.vertex_available(),
+ "bigquery": bq.active(),
+ "judge_model": config.JUDGE_MODEL,
+ "vertex_location": config.VERTEX_LOCATION,
+ },
}
@@ -215,7 +223,9 @@ def add_note(req: NoteReq):
@app.get("/api/corpus/search")
def corpus_search(clause_type: str = "liability_cap", limit: int = 5):
- return {"results": db.corpus_search(clause_type, limit)}
+ # Prefer BigQuery when GCP is configured, fall back to SQLite
+ results = bq.search(clause_type, limit) or db.corpus_search(clause_type, limit)
+ return {"results": results, "source": "bigquery" if bq.active() else "sqlite"}
@app.post("/api/corpus/build/edgar")
diff --git a/backend/requirements.txt b/backend/requirements.txt
new file mode 100644
index 0000000..4a8ead4
--- /dev/null
+++ b/backend/requirements.txt
@@ -0,0 +1,9 @@
+fastapi==0.115.6
+uvicorn[standard]==0.34.0
+python-dotenv==1.0.1
+httpx==0.28.1
+openai==1.59.6
+pydantic==2.10.4
+# GCP — Vertex AI + BigQuery (required for Cloud Run / GCP prize track)
+google-cloud-aiplatform>=1.60.0
+google-cloud-bigquery>=3.11.0