Add plan for voice-enabled knowledge agent

agent-bot-plan.md

@@ -0,0 +1,72 @@
+# Agent Bot Plan: PDF/Website Knowledge + Voice Conversation
+
+## Goal
+Build a personal assistant that learns from PDFs or websites and holds voice conversations with answers grounded in the source repository content.
+
+## Core Capabilities
+- **Ingest sources:** Upload PDF(s) or provide website URLs.
+- **Index knowledge:** Extract text, chunk it, and store embeddings in a vector database.
+- **Grounded responses:** Use retrieval-augmented generation (RAG) to answer only from indexed sources.
+- **Voice interface:** Speech-to-text (STT) for input and text-to-speech (TTS) for responses.
+- **Conversation memory:** Maintain short-term context while citing or summarizing sources.
+
+## Suggested Architecture
+### 1) Ingestion & Processing
+- **PDF pipeline:**
+  - Extract text from PDF.
+  - Normalize whitespace, remove headers/footers if needed.
+  - Chunk by sections or fixed token size (e.g., 500–1,000 tokens with overlap).
+- **Website pipeline:**
+  - Fetch content, remove boilerplate, extract readable text.
+  - Respect robots.txt and rate limits.
+  - Chunk similarly to PDF pipeline.
+
+### 2) Embeddings & Vector Store
+- Generate embeddings for each chunk.
+- Store embeddings and metadata (source URL/file, page, section) in a vector database.
+- When asked a question, retrieve the most relevant chunks.
+
+### 3) Retrieval-Augmented Generation (RAG)
+- Combine retrieved chunks with the user’s question.
+- Use a system prompt that **restricts answers to retrieved sources**.
+- Optionally provide inline citations from chunk metadata.
+
+### 4) Voice Conversation Layer
+- **STT:** Convert user audio to text.
+- **TTS:** Convert model output to audio.
+- **Streaming:** Stream audio responses for a more conversational experience.
+
+## Recommended Tech Stack (Example)
+- **Backend:** Python + FastAPI
+- **Embeddings:** OpenAI or local embeddings (e.g., sentence-transformers)
+- **Vector DB:** FAISS (local), Chroma, or Pinecone
+- **RAG Orchestration:** Lightweight custom pipeline or frameworks like LangChain/LlamaIndex
+- **STT:** OpenAI Whisper or Deepgram
+- **TTS:** ElevenLabs, OpenAI TTS, or Coqui TTS
+- **Frontend:** Minimal web app with microphone input + audio output
+
+## Data Flow
+1. User uploads PDF or enters website URL.
+2. Server extracts, chunks, and embeds content.
+3. Vector DB stores embeddings + metadata.
+4. User asks a question by voice.
+5. STT converts voice to text.
+6. RAG retrieves relevant chunks and generates a grounded response.
+7. TTS converts the answer to speech and streams it back.
+
+## Safety & Quality
+- **Grounding:** Never answer without relevant chunks; ask follow-up questions if retrieval is weak.
+- **Source citations:** Provide page numbers or URLs where possible.
+- **Privacy:** Store user data locally or encrypt at rest if hosted.
+
+## MVP Milestones
+1. **Text-only RAG demo** (PDF + website ingestion, Q&A in text).
+2. **Add STT + TTS** for voice input/output.
+3. **UI polish** (conversation history, source citations, upload status).
+4. **Performance tuning** (better chunking, caching, retrieval filters).
+
+## Stretch Goals
+- Multi-document collections with tags and filters.
+- User-specific memory (preferences, frequently asked topics).
+- Offline mode (local embeddings + local TTS/STT).
+- Fine-grained permissions and access control.