feat: Add BYOK (Bring Your Own Key) support for Groq API

backend/app/.gitignore

@@ -0,0 +1 @@
+.env

backend/app/logging/logging_config.py

@@ -24,8 +24,12 @@ def setup_logger(name: str) -> logging.Logger:
         datefmt="%Y-%m-%d %H:%M:%S"
     )
 
-    # Console Handler
-    console_handler = logging.StreamHandler(sys.stdout)
+    # Console Handler (force UTF-8 to prevent UnicodeEncodeError on Windows with emoji)
+    try:
+        utf8_stream = open(sys.stdout.fileno(), mode="w", encoding="utf-8", closefd=False)
+        console_handler = logging.StreamHandler(utf8_stream)
+    except Exception:
+        console_handler = logging.StreamHandler(sys.stdout)
     console_handler.setLevel(logging.INFO)
     console_handler.setFormatter(formatter)
     logger.addHandler(console_handler)

backend/app/modules/bias_detection/check_bias.py

@@ -32,10 +32,9 @@
 
 load_dotenv()
 
-client = Groq(api_key=os.getenv("GROQ_API_KEY"))
 
-
-def check_bias(text):
+def check_bias(text, api_key: str, groq_model: str = "llama-3.3-70b-versatile"):
+    client = Groq(api_key=api_key)
     try:
         logger.debug(f"Raw article text: {text}")
         logger.debug(f"JSON dump of text: {json.dumps(text)}")
@@ -61,7 +60,7 @@ def check_bias(text):
                     "content": (f"Give bias score to the following article \n\n{text}"),
                 },
             ],
-            model="gemma2-9b-it",
+            model=groq_model,
             temperature=0.3,
             max_tokens=512,
         )

backend/app/modules/chat/llm_processing.py

@@ -32,8 +32,6 @@
 
 load_dotenv()
 
-client = Groq(api_key=os.getenv("GROQ_API_KEY"))
-
 
 def build_context(docs):
     return "\n".join(
@@ -42,10 +40,19 @@ def build_context(docs):
     )
 
 
-def ask_llm(question, docs):
-    context = build_context(docs)
-    logger.debug(f"Generated context for LLM:\n{context}")
-    prompt = f"""You are an assistant that answers based on context.
+def ask_llm(question, docs, api_key: str, groq_model: str = "llama-3.3-70b-versatile", article_text: str = ""):
+    client = Groq(api_key=api_key)
+    pinecone_context = build_context(docs)
+    logger.debug(f"Generated context for LLM:\n{pinecone_context}")
+
+    context_parts = []
+    if article_text:
+        context_parts.append(f"=== Full Article ===\n{article_text}")
+    if pinecone_context:
+        context_parts.append(f"=== Fact-Check Notes ===\n{pinecone_context}")
+    context = "\n\n".join(context_parts) or "No context available."
+
+    prompt = f"""You are an assistant that answers questions about a news article.
 
 Context:
 {context}
@@ -55,7 +62,7 @@ def ask_llm(question, docs):
 """
 
     response = client.chat.completions.create(
-        model="gemma2-9b-it",
+        model=groq_model,
         messages=[
             {"role": "system", "content": "Use only the context to answer."},
             {"role": "user", "content": prompt},

backend/app/modules/facts_check/llm_processing.py

@@ -34,10 +34,9 @@
 
 load_dotenv()
 
-client = Groq(api_key=os.getenv("GROQ_API_KEY"))
-
 
 def run_claim_extractor_sdk(state):
+    client = Groq(api_key=state["groq_api_key"])
     try:
         text = state.get("cleaned_text")
         if not text:
@@ -63,7 +62,7 @@ def run_claim_extractor_sdk(state):
                     ),
                 },
             ],
-            model="gemma2-9b-it",
+            model=state.get("groq_model", "llama-3.3-70b-versatile"),
             temperature=0.3,
             max_tokens=512,
         )
@@ -87,7 +86,8 @@ def run_claim_extractor_sdk(state):
         }
 
 
-def run_fact_verifier_sdk(search_results):
+def run_fact_verifier_sdk(search_results, api_key: str, groq_model: str = "llama-3.3-70b-versatile"):
+    client = Groq(api_key=api_key)
     try:
         results_list = []
 
@@ -128,7 +128,7 @@ def run_fact_verifier_sdk(search_results):
                         ),
                     },
                 ],
-                model="gemma2-9b-it",
+                model=groq_model,
                 temperature=0.3,
                 max_tokens=256,
             )

backend/app/modules/facts_check/web_search.py

@@ -32,6 +32,8 @@ def search_google(query):
         f"https://www.googleapis.com/customsearch/v1?key={GOOGLE_SEARCH}&cx=f637ab77b5d8b4a3c&q={query}"
     )
     res = results.json()
+    if "items" not in res:
+        return []
     first = {}
     first["title"] = res["items"][0]["title"]
     first["link"] = res["items"][0]["link"]

backend/app/modules/langgraph_builder.py

@@ -52,6 +52,8 @@ class MyState(TypedDict):
     score: int
     retries: int
     status: str
+    groq_api_key: str
+    groq_model: str
 
 
 def build_langgraph():

backend/app/modules/langgraph_nodes/fact_check.py

@@ -32,12 +32,10 @@ def run_fact_check(state):
         verifications, error_message = run_fact_check_pipeline(state)
 
         if error_message:
-            logger.error(f"Error in fact-checking: {error_message}")
-            return {
-                "status": "error",
-                "error_from": "fact_checking",
-                "message": f"{error_message}",
-            }
+            # Soft failure — web search quota/key issue. Continue with empty facts
+            # so the rest of the pipeline (generate_perspective, store_and_send) still runs.
+            logger.warning(f"Fact-checking skipped (non-fatal): {error_message}")
+            verifications = []
 
     except Exception as e:
         logger.exception(f"Unexpected error in fact-checking: {e}")

backend/app/modules/langgraph_nodes/generate_perspective.py

@@ -36,15 +36,6 @@ class PerspectiveOutput(BaseModel):
     perspective: str = Field(..., description="Generated opposite perspective")
 
 
-my_llm = "llama-3.3-70b-versatile"
-
-llm = ChatGroq(model=my_llm, temperature=0.7)
-
-structured_llm = llm.with_structured_output(PerspectiveOutput)
-
-
-chain = prompt | structured_llm
-
 
 def generate_perspective(state):
     try:
@@ -56,17 +47,19 @@ def generate_perspective(state):
 
         if not text:
             raise ValueError("Missing or empty 'cleaned_text' in state")
-        elif not facts:
-            raise ValueError("Missing or empty 'facts' in state")
 
+        llm = ChatGroq(model=state["groq_model"], temperature=0.7, api_key=state["groq_api_key"])
+        chain = prompt | llm.with_structured_output(PerspectiveOutput)
+
+        # facts may be empty if web search failed — generate perspective without them
         facts_str = "\n".join(
             [
                 f"Claim: {f['original_claim']}\n"
                 "Verdict: {f['verdict']}\nExplanation: "
                 "{f['explanation']}"
-                for f in state["facts"]
+                for f in (facts or [])
             ]
-        )
+        ) or "No verified facts available."
 
         result = chain.invoke(
             {

backend/app/modules/langgraph_nodes/judge.py

@@ -23,14 +23,13 @@
 logger = setup_logger(__name__)
 
 # Init once
-groq_llm = ChatGroq(
-    model="gemma2-9b-it",
-    temperature=0.0,
-    max_tokens=10,
-)
-
-
 def judge_perspective(state):
+    groq_llm = ChatGroq(
+        model=state.get("groq_model", "llama-3.3-70b-versatile"),
+        temperature=0.0,
+        max_tokens=10,
+        api_key=state["groq_api_key"],
+    )
     try:
         perspective_obj = state.get("perspective")
         text = getattr(perspective_obj, "perspective", "").strip()

backend/app/modules/langgraph_nodes/sentiment.py

@@ -23,10 +23,9 @@
 
 load_dotenv()
 
-client = Groq(api_key=os.getenv("GROQ_API_KEY"))
-
 
 def run_sentiment_sdk(state):
+    client = Groq(api_key=state["groq_api_key"])
     try:
         text = state.get("cleaned_text")
         if not text:
@@ -49,7 +48,7 @@ def run_sentiment_sdk(state):
                     ),
                 },
             ],
-            model="gemma2-9b-it",
+            model=state.get("groq_model", "llama-3.3-70b-versatile"),
             temperature=0.2,
             max_tokens=3,
         )

backend/app/modules/pipeline.py

@@ -64,8 +64,8 @@ def run_scraper_pipeline(url: str) -> dict:
     return result
 
 
-def run_langgraph_workflow(state: dict):
+def run_langgraph_workflow(state: dict, api_key: str, groq_model: str = "llama-3.3-70b-versatile"):
     """Execute the pre-compiled LangGraph workflow."""
-    result = _LANGGRAPH_WORKFLOW.invoke(state)
+    result = _LANGGRAPH_WORKFLOW.invoke({**state, "groq_api_key": api_key, "groq_model": groq_model})
     logger.info("LangGraph workflow executed successfully.")
     return result

backend/app/routes/routes.py

@@ -30,7 +30,7 @@
 """
 
 
-from fastapi import APIRouter
+from fastapi import APIRouter, Request, HTTPException
 from pydantic import BaseModel
 from app.modules.pipeline import run_scraper_pipeline
 from app.modules.pipeline import run_langgraph_workflow
@@ -52,6 +52,7 @@ class URlRequest(BaseModel):
 
 class ChatQuery(BaseModel):
     message: str
+    article_text: str = ""
 
 
 @router.get("/")
@@ -60,26 +61,37 @@ async def home():
 
 
 @router.post("/bias")
-async def bias_detection(request: URlRequest):
-    content = await asyncio.to_thread(run_scraper_pipeline, (request.url))
-    bias_score = await asyncio.to_thread(check_bias, (content))
+async def bias_detection(url_request: URlRequest, request: Request):
+    api_key = request.headers.get("x-byok-api-key")
+    if not api_key:
+        raise HTTPException(status_code=401, detail="Missing X-BYOK-Api-Key header")
+    groq_model = request.headers.get("x-byok-model", "llama-3.3-70b-versatile")
+    content = await asyncio.to_thread(run_scraper_pipeline, url_request.url)
+    bias_score = await asyncio.to_thread(check_bias, content, api_key, groq_model)
     logger.info(f"Bias detection result: {bias_score}")
     return bias_score
 
 
 @router.post("/process")
-async def run_pipelines(request: URlRequest):
-    article_text = await asyncio.to_thread(run_scraper_pipeline, (request.url))
+async def run_pipelines(url_request: URlRequest, request: Request):
+    api_key = request.headers.get("x-byok-api-key")
+    if not api_key:
+        raise HTTPException(status_code=401, detail="Missing X-BYOK-Api-Key header")
+    groq_model = request.headers.get("x-byok-model", "llama-3.3-70b-versatile")
+    article_text = await asyncio.to_thread(run_scraper_pipeline, url_request.url)
     logger.debug(f"Scraper output: {json.dumps(article_text, indent=2, ensure_ascii=False)}")
-    data = await asyncio.to_thread(run_langgraph_workflow, (article_text))
+    data = await asyncio.to_thread(run_langgraph_workflow, article_text, api_key, groq_model)
     return data
 
 
 @router.post("/chat")
-async def answer_query(request: ChatQuery):
-    query = request.message
+async def answer_query(chat_request: ChatQuery, request: Request):
+    api_key = request.headers.get("x-byok-api-key")
+    if not api_key:
+        raise HTTPException(status_code=401, detail="Missing X-BYOK-Api-Key header")
+    groq_model = request.headers.get("x-byok-model", "llama-3.3-70b-versatile")
+    query = chat_request.message
     results = search_pinecone(query)
-    answer = ask_llm(query, results)
+    answer = ask_llm(query, results, api_key, groq_model, chat_request.article_text)
     logger.info(f"Chat answer generated: {answer}")
-
     return {"answer": answer}

backend/app/utils/fact_check_utils.py

@@ -76,5 +76,5 @@ def run_fact_check_pipeline(state):
         return [], "All claim searches failed or returned no results."
 
     # Step 3: Verify facts using LLM
-    final = run_fact_verifier_sdk(search_results)
+    final = run_fact_verifier_sdk(search_results, state["groq_api_key"], state.get("groq_model", "llama-3.3-70b-versatile"))
     return final.get("verifications", []), None

backend/pyproject.toml

@@ -13,6 +13,7 @@ dependencies = [
     "groq>=0.28.0",
     "langchain>=0.3.25",
     "langchain-community>=0.3.25",
+    "langchain-google-genai>=2.1.12",
     "langchain-groq>=0.3.2",
     "langgraph>=0.4.8",
     "logging>=0.4.9.6",