Use DeepSeek V4 Flash with bilingual query planning

DEPLOYMENT.md

@@ -5,23 +5,24 @@ Target Space: `https://huggingface.co/spaces/wingsxyp/algorithm-research-deck`
 ## Token roles
 
 - **Write token**: use only on your computer for Git authentication and pushing files to the Space repository.
-- **Read token**: add to the Space as the Secret `HF_TOKEN`; the running application uses it to download the configured open model.
+- **Inference token**: add to the Space as the Secret `HF_TOKEN`. It needs permission to call Hugging Face Inference Providers, but does not need repository write access.
 - Never commit either token. Never put the write token in Space Secrets.
 
 ## Space settings
 
-The repository README configures a Gradio Space. In the Space settings, select ZeroGPU hardware if it is not already selected.
+The repository README configures a Gradio Space. DeepSeek-V4-Flash is called through Hugging Face Inference Providers because its 284B total parameters are too large for direct loading in one ZeroGPU allocation. The Space can remain on ZeroGPU, although the application itself does not request ZeroGPU for model inference.
 
 Add this Secret:
 
 ```text
-HF_TOKEN=<your read token>
+HF_TOKEN=<your fine-grained inference token>
 ```
 
 Add these Variables:
 
 ```text
-HF_MODEL_ID=Qwen/Qwen2.5-7B-Instruct
+HF_MODEL_ID=deepseek-ai/DeepSeek-V4-Flash
+HF_PROVIDER=auto
 CONTACT_EMAIL=<your email for polite arXiv requests>
 ```
 
@@ -56,9 +57,9 @@ Open the Space build logs. A successful build installs Gradio, Spaces, Transform
 Verify:
 
 - arXiv search completes
-- ZeroGPU allocation starts for model synthesis
+- DeepSeek-V4-Flash responds through an Inference Provider
 - the presentation appears in the right panel
 - HTML, JSON, and ZIP downloads work
 - the downloaded HTML opens independently in a browser
 
-If model loading fails, first confirm `HF_MODEL_ID`, the read token's model access, and ZeroGPU quota. Do not replace the read token with the write token.
+If inference fails, first confirm `HF_MODEL_ID`, the token's Inference Providers permission, provider availability, and account credits. Do not replace the inference token with the repository write token.

README.md

@@ -14,17 +14,18 @@ license: apache-2.0
 
 # Algorithm Research Deck
 
-Turn an algorithm research topic into an interactive, paper-backed HTML presentation. The model first expands the topic into historical, milestone, current, and benchmark-oriented search queries. The Space searches arXiv with OpenAlex fallback, uses ZeroGPU for open-model synthesis, previews the result, and provides HTML, JSON, and ZIP downloads.
+Turn an algorithm research topic into an interactive, paper-backed HTML presentation. DeepSeek-V4-Flash first converts Chinese inputs into a precise English technical interpretation and expands them into historical, milestone, current, and benchmark-oriented search queries. The Space searches arXiv with OpenAlex fallback, uses Hugging Face Inference Providers for synthesis, previews the result, and provides HTML, JSON, and ZIP downloads.
 
 ## Space configuration
 
 Secret:
 
-- `HF_TOKEN`: a read token used only to download the configured model. Public models may not require it, but setting it avoids anonymous download limits.
+- `HF_TOKEN`: a fine-grained token with permission to call Hugging Face Inference Providers. Repository write permission is neither needed nor recommended.
 
 Variables:
 
-- `HF_MODEL_ID`: defaults to `Qwen/Qwen2.5-7B-Instruct`
+- `HF_MODEL_ID`: defaults to `deepseek-ai/DeepSeek-V4-Flash`
+- `HF_PROVIDER`: defaults to `auto`; optionally choose a currently live provider
 - `CONTACT_EMAIL`: contact address included in polite arXiv requests
 
 The token used on your computer to push this repository must be a write token. Never add that write token to Space Secrets.

app.py

@@ -16,20 +16,17 @@ from xml.etree import ElementTree
 
 import gradio as gr
 import httpx
-import spaces
-import torch
+from huggingface_hub import InferenceClient
 from pydantic import BaseModel, Field
 from pypdf import PdfReader
-from transformers import AutoModelForCausalLM, AutoTokenizer
 
 
 BASE_DIR = Path(__file__).resolve().parent
 STATIC_DIR = BASE_DIR / "static"
 DATA_DIR = Path("/data/research-decks") if Path("/data").exists() else BASE_DIR / "output"
 DATA_DIR.mkdir(parents=True, exist_ok=True)
-MODEL_ID = os.getenv("HF_MODEL_ID", "Qwen/Qwen2.5-7B-Instruct")
-_model = None
-_tokenizer = None
+MODEL_ID = os.getenv("HF_MODEL_ID", "deepseek-ai/DeepSeek-V4-Flash")
+HF_PROVIDER = os.getenv("HF_PROVIDER", "auto")
 
 
 class ResearchRequest(BaseModel):
@@ -184,14 +181,18 @@ def deduplicate(papers: list[dict[str, Any]]) -> list[dict[str, Any]]:
     return output
 
 
-def plan_search_queries(request: ResearchRequest) -> list[str]:
+def plan_search_queries(request: ResearchRequest) -> tuple[list[str], dict[str, Any]]:
     prompt = json.dumps({
         "topic": request.topic,
         "purpose": request.purpose,
+        "audience": request.audience,
         "focus": request.focus,
         "date_range": request.date_range,
+        "language": request.language,
         "instruction": (
-            "Understand the technical topic before searching. Return JSON with search_queries containing 3 to 5 concise English academic queries. "
+            "First translate and normalize every user field into precise English technical meaning without dropping the original intent. "
+            "Return english_understanding with topic, purpose, audience, focus, date_range, key_concepts, ambiguities, and scope. "
+            "Also return search_queries containing 3 to 5 concise English academic queries. "
             "Cover historical/classical terminology, milestone method families, current terminology, and one limitation or benchmark angle. "
             "Queries must be suitable for arXiv and OpenAlex. Do not include site operators."
         ),
@@ -202,9 +203,11 @@ def plan_search_queries(request: ResearchRequest) -> list[str]:
         650,
     )
     queries = [str(x).strip() for x in result.get("search_queries", []) if str(x).strip()]
-    if request.topic not in queries:
-        queries.append(request.topic)
-    return queries[:6]
+    english_understanding = result.get("english_understanding") or {}
+    fallback_query = str(english_understanding.get("topic") or request.topic).strip()
+    if fallback_query and fallback_query not in queries:
+        queries.append(fallback_query)
+    return queries[:6], english_understanding
 
 
 def search_literature(queries: list[str], total_limit: int) -> tuple[list[dict[str, Any]], list[str]]:
@@ -228,30 +231,18 @@ def search_literature(queries: list[str], total_limit: int) -> tuple[list[dict[s
     return deduplicate(papers)[:total_limit], warnings
 
 
-@spaces.GPU(duration=120)
 def generate_text(system: str, prompt: str, max_tokens: int) -> str:
-    global _model, _tokenizer
-    token = os.getenv("HF_TOKEN") or None
-    if _tokenizer is None:
-        _tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, token=token)
-    if _model is None:
-        _model = AutoModelForCausalLM.from_pretrained(
-            MODEL_ID,
-            token=token,
-            torch_dtype=torch.bfloat16,
-            device_map="auto",
-        )
-    messages = [{"role": "system", "content": system}, {"role": "user", "content": prompt}]
-    model_input = _tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-    inputs = _tokenizer(model_input, return_tensors="pt").to(_model.device)
-    with torch.inference_mode():
-        generated = _model.generate(
-            **inputs,
-            max_new_tokens=max_tokens,
-            do_sample=False,
-            repetition_penalty=1.05,
-        )
-    return _tokenizer.decode(generated[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
+    token = os.getenv("HF_TOKEN")
+    if not token:
+        raise RuntimeError("Space Secret HF_TOKEN is missing")
+    client = InferenceClient(provider=HF_PROVIDER, api_key=token, timeout=180)
+    response = client.chat.completions.create(
+        model=MODEL_ID,
+        messages=[{"role": "system", "content": system}, {"role": "user", "content": prompt}],
+        temperature=0.2,
+        max_tokens=max_tokens,
+    )
+    return response.choices[0].message.content
 
 
 def chat_json(system: str, prompt: str, max_tokens: int = 6000) -> dict[str, Any]:
@@ -298,7 +289,12 @@ def extract_pdf(paper: dict[str, Any], max_chars: int = 14000) -> str:
     return "\n".join(pages)[:max_chars]
 
 
-def build_report(request: ResearchRequest, papers: list[dict[str, Any]], deep_texts: dict[str, str]) -> dict[str, Any]:
+def build_report(
+    request: ResearchRequest,
+    papers: list[dict[str, Any]],
+    deep_texts: dict[str, str],
+    english_understanding: dict[str, Any],
+) -> dict[str, Any]:
     evidence = []
     for paper in papers:
         item = {k: paper.get(k) for k in ("id", "title", "summary", "authors", "published", "updated", "url")}
@@ -307,6 +303,7 @@ def build_report(request: ResearchRequest, papers: list[dict[str, Any]], deep_te
         evidence.append(item)
     payload = {
         "request": request.model_dump(),
+        "english_understanding": english_understanding,
         "research_cutoff": datetime.now(timezone.utc).date().isoformat(),
         "evidence": evidence,
         "schema": {
@@ -417,7 +414,7 @@ def run_research_ui(
     )
     budget = DEPTHS[request.depth]
     progress(0.05, desc="模型正在理解主题并规划检索词")
-    queries = plan_search_queries(request)
+    queries, english_understanding = plan_search_queries(request)
     progress(0.16, desc="正在检索 arXiv 和 OpenAlex")
     user_papers = normalize_user_papers(request.paper_urls)
     discovered_papers, search_warnings = search_literature(queries, budget["discover"])
@@ -427,7 +424,7 @@ def run_research_ui(
     if not papers:
         raise gr.Error("没有检索到论文，请尝试英文关键词或提供 arXiv 链接。")
     screen = papers[: budget["screen"]]
-    progress(0.28, desc="ZeroGPU 正在筛选关键论文")
+    progress(0.28, desc="DeepSeek-V4-Flash 正在筛选关键论文")
     selected_ids = select_papers(request, screen, budget["screen"], budget["deep"])
     if not selected_ids:
         selected_ids = [p["id"] for p in screen[: budget["deep"]]]
@@ -440,8 +437,8 @@ def run_research_ui(
             except Exception as exc:
                 deep_texts[paper["id"]] = f"PDF extraction failed: {exc}"
         progress(0.48 + 0.16 * (index + 1) / len(screen), desc="正在提取论文证据")
-    progress(0.68, desc="ZeroGPU 正在生成技术路线")
-    report = build_report(request, screen, deep_texts)
+    progress(0.68, desc="DeepSeek-V4-Flash 正在生成技术路线")
+    report = build_report(request, screen, deep_texts, english_understanding)
     cutoff = datetime.now(timezone.utc).date().isoformat()
     content = render_presentation(report, screen, cutoff)
     job_dir = DATA_DIR / uuid.uuid4().hex

requirements.txt

@@ -1,4 +1,2 @@
 httpx==0.28.1
 pypdf==5.5.0
-transformers
-accelerate==1.6.0