feat: change model to deepseek

.env.example

@@ -13,7 +13,7 @@ APP_PASSWORD=change-me
 # Default admin passcode is "password"; override this for real deployments.
 # ADMIN_PASSCODE=change-me
 
-LLM_MODEL=Qwen/Qwen2.5-7B-Instruct
+LLM_MODEL=deepseek-ai/DeepSeek-V4-Flash
 EMBED_MODEL=BAAI/bge-small-en-v1.5
 
 # Leave reasoning hidden in the student UI. Set true only when debugging models

README.md

@@ -18,7 +18,7 @@ pinned: false
    - `APP_USERNAME` — defaults to `student`.
    - `DATASET_ID` — defaults to `<SPACE_ID>-corpus` on Spaces and is created automatically when needed.
    - `ADMIN_PASSCODE` — defaults to `password`; set this for real deployments.
-   - *(optional)* `LLM_MODEL` — defaults to `Qwen/Qwen2.5-7B-Instruct`.
+   - *(optional)* `LLM_MODEL` — defaults to `deepseek-ai/DeepSeek-V4-Flash`; change this Space variable to switch models without a code push.
    - *(optional)* `EMBED_MODEL` — defaults to `BAAI/bge-small-en-v1.5`.
 4. Restart the Space. Open `/admin`, sign in with the passcode, and upload rubric files (`.pdf`, `.docx`, `.txt`, `.md`).
 5. Share the root URL with students. They can paste or upload tutor feedback and upload coursework locally for the chat session.
@@ -92,4 +92,4 @@ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLI
 
 If you use or adapt this tool in your work, please cite:
 
-Abbas, N. (under review). Investigating Student Perceptions of an AI-Powered Chatbot to Support Feedback Interpretation and Uptake in Higher Education. Manuscript submitted to Assessment and Evaluation in Higher Education.
+Abbas, N. (under review). Investigating Student Perceptions of an AI-Powered Chatbot to Support Feedback Interpretation and Uptake in Higher Education. Manuscript submitted to Assessment and Evaluation in Higher Education.

backend/config.py

@@ -30,8 +30,10 @@ HF_TOKEN = os.environ.get("HF_TOKEN")
 DATASET_ID = os.environ.get("DATASET_ID") or _default_dataset_id()
 ADMIN_PASSCODE = os.environ.get("ADMIN_PASSCODE", "password")
 
-LLM_MODEL = os.environ.get("LLM_MODEL", "Qwen/Qwen2.5-7B-Instruct")
+DEFAULT_LLM_MODEL = "deepseek-ai/DeepSeek-V4-Flash"
+LLM_MODEL = os.environ.get("LLM_MODEL", DEFAULT_LLM_MODEL).strip() or DEFAULT_LLM_MODEL
 LLM_MODEL_LOWER = LLM_MODEL.lower()
+LLM_PROVIDER = os.environ.get("LLM_PROVIDER", "").strip() or None
 EMBED_MODEL = os.environ.get("EMBED_MODEL", "BAAI/bge-small-en-v1.5")
 LLM_EXTRA_BODY_JSON = os.environ.get("LLM_EXTRA_BODY_JSON", "").strip()
 LLM_FINAL_ANSWER_EXTRA_BODY_JSON = os.environ.get(
@@ -55,4 +57,3 @@ BACKEND_INTERNAL_URL = os.environ.get("BACKEND_INTERNAL_URL", f"http://127.0.0.1
 TMP_UPLOAD_DIR = Path(os.environ.get("TMP_UPLOAD_DIR", str(REPO_ROOT / ".tmp_uploads")))
 LANCEDB_DIR = Path(os.environ.get("LANCEDB_PATH", str(REPO_ROOT / ".lancedb")))
 LANCEDB_PATH = str(LANCEDB_DIR)
-

backend/main.py

@@ -45,6 +45,7 @@ TMP_UPLOAD_DIR = config.TMP_UPLOAD_DIR
 LANCEDB_DIR = config.LANCEDB_DIR
 LLM_MODEL = config.LLM_MODEL
 LLM_MODEL_LOWER = config.LLM_MODEL_LOWER
+LLM_PROVIDER = config.LLM_PROVIDER
 EMBED_MODEL = config.EMBED_MODEL
 HF_TOKEN = config.HF_TOKEN
 LLM_EXTRA_BODY_JSON = config.LLM_EXTRA_BODY_JSON
@@ -193,7 +194,10 @@ async def lifespan(app: FastAPI):  # noqa: ARG001
             logger.exception("Vector cache restore failed: %s", exc)
 
     rag = RagEngine()
-    llm_client = AsyncInferenceClient(model=LLM_MODEL, token=HF_TOKEN)
+    llm_client_kwargs: dict[str, Any] = {"model": LLM_MODEL, "token": HF_TOKEN}
+    if LLM_PROVIDER:
+        llm_client_kwargs["provider"] = LLM_PROVIDER
+    llm_client = AsyncInferenceClient(**llm_client_kwargs)
 
     personas_store.load()
 
@@ -238,6 +242,7 @@ async def health() -> dict:
     return {
         "ok": True,
         "model": LLM_MODEL,
+        "provider": LLM_PROVIDER or "auto",
         "embed_model": EMBED_MODEL,
         "dataset_id": config.DATASET_ID,
         "admin_passcode_defaulted": config.ADMIN_PASSCODE == "password" and "ADMIN_PASSCODE" not in os.environ,