Streamlit RAG chat app with Pinecone, SambaNova LLM, and Supabase-backed auth.

• Upload PDF, DOCX, TXT files. Text is extracted, chunked, embedded with BAAI/bge-large-en-v1.5, and indexed in Pinecone.
• Ask questions in chat. If web search is enabled, the app uses web results; else it tries RAG over Pinecone (if your docs are indexed); default it answers directly with the LLM using conversation facts.
• Users sign up and log in. Passwords are hashed and profiles plus chats are stored in Supabase.
• Sidebar controls: API keys, web-search toggle, document processing, chat search, and new chat.

flowchart LR U[User: Streamlit UI] -->|Login/Signup| SB[(Supabase profiles)] U -->|Upload PDFs| PDF[PyMuPDF] PDF --> SPLIT[TextSplitter 512/128] SPLIT --> EMB[HF Embeddings bge-large] EMB --> IDX[(Pinecone Index<br/>namespace = username)] U -->|Ask| Q[Question] Q --> WEB{Web search enabled?} WEB -- Yes --> SERP[Serper k=5] --> WSUM[LLM summarize] WEB -- No --> RET{Docs indexed?} RET -- Yes --> R[VDB Retriever k=10, MMR] --> COMB[LLM combine] RET -- No --> DIRECT[Direct LLM with conversation facts] WSUM --> OUT[Answer+Sources] COMB --> OUT DIRECT --> OUT OUT --> SB OUT --> U 

• Streamlit UI: chat interface, sidebar settings, chat list and search.
• Auth + Persistence: Supabase table profiles keeps username, password (SHA‑256 hash), and a list of chats. Chats store id, title, messages, created_at, updated_at.
• File ingestion: PDF via PyMuPDF; TXT via UTF-8 decode; DOCX via python-docx. Uses RecursiveCharacterTextSplitter with chunk_size=512, chunk_overlap=128.
• Embeddings: HuggingFaceEmbeddings("BAAI/bge-large-en-v1.5").
• Vector store: Pinecone via PINECONE_INDEX_NAME and PINECONE_API_KEY. Namespace per user = username. IDs are SHA‑1 of content + source.
• LLM: SambaNovaCloud with selectable models in the sidebar (default: DeepSeek-V3.1).
• Answering flow:
  1. If web search is toggled and Serper key is provided, use web results with LLM summarization.
  2. Else, if your docs are indexed, retrieve with Pinecone (k=10, MMR) and answer from context.
  3. Else, answer directly with the LLM, using conversation facts from memory.
• Guardrails: Detoxify-based input screening flags toxic content and adds a warning to the assistant response.

• Python 3.10+
• Accounts/keys:
  • Supabase: SUPABASE_URL, SUPABASE_KEY
  • Pinecone: PINECONE_API_KEY, PINECONE_INDEX_NAME
  • SambaNova API key (entered in sidebar for answering). For conversation summarization memory, set environment variable SAMBANOVA_API_KEY.
  • Serper API key for optional web search (entered in sidebar)
  • Username constraint: must be a Gmail address (e.g., name@gmail.com).

pip install -r requirements.txt

If you use GPU embeddings or other models, install the matching extras separately.

Create a .env that the app loads. By default the code calls in config.py:

load_dotenv(dotenv_path=".env")

Options:

• Put your variables in .env, or
• Change the dotenv_path in config.py to your desired location.

Minimum variables in .env file:

SUPABASE_URL=... SUPABASE_KEY=... PINECONE_API_KEY=... PINECONE_INDEX_NAME=... SAMBANOVA_API_KEY=... 

streamlit run app.py

In the sidebar:

1. Paste SambaNova API key. Optionally paste Serper API key.
2. Toggle Web search if you want web fallback.
3. Upload PDFs and click Process Documents.
4. Start chatting. Use Chats to create or open a conversation. Use Search chats to filter by message text.

• Chunking: 512 characters, 128 overlap.
• Retrieval: MMR with k=10.
• Pinecone namespace: the current username.
• Password hashing: SHA‑256 for demo. For production, prefer a dedicated auth provider or salted password hashing (e.g., Argon2/bcrypt) with rate‑limiting and MFA.
• If Serper key is missing or web search is off, the app falls back to RAG or direct LLM.
• Timezone: Asia/Kolkata by default. Change TIMEZONE in config.py if needed.

• Pinecone connection: ensure PINECONE_INDEX_NAME and PINECONE_API_KEY are set. The app uses the index name (not host).
• “Supabase credentials not found”: set SUPABASE_URL and SUPABASE_KEY in the .env that config.py loads.
• No answer from docs: confirm documents were processed, embeddings created, and the correct namespace is used.
• Serper missing: web search button can be ON, but without a Serper key the code will skip web and use LLM only.
• SambaNova summarization: conversation summary memory uses SAMBANOVA_API_KEY from environment. Set it if summaries are not updating.

MIT. See LICENSE.