Sanki - The Vector Bot

Overview

I built Sanki as a Telegram-first personal AI that works as an extension of my own context rather than a generic chatbot. It accepts text or voice, reasons through a Gemini-powered agent, selectively queries a personal vector database, and returns grounded responses in seconds.

What It Delivers

• Text and voice interaction through Telegram, with the architecture adaptable to additional messaging platforms.
• Short-term conversational context combined with long-term semantic memory.
• A vector database containing more than 100,000 words from essays, projects, background material, and thinking patterns.
• Meaning-based retrieval through embeddings and RAG rather than simple keyword lookup.
• Selective memory access only when relevant to the user's request.
• Tool access spanning email, calendars, tasks, CRM, image generation, and document summarization.

What I Did

• Designed the assistant architecture, memory hierarchy, vector-retrieval logic, system prompt, messaging flow, and tool integrations.
• Prepared and embedded more than 100,000 words of personal context using Gemini embeddings.
• Defined the distinction between short-term chat memory and long-term personal memory.
• Built the system before persistent personal memory became a common default feature in mainstream assistants.
• Structured the agent to retrieve personal context only when required rather than flooding every prompt.

Outcomes

• A personalized AI capable of reasoning from the user's actual history and work.
• Semantic retrieval grounded in personal documents and projects.
• A single conversational interface capable of accessing multiple operational tools.
• A scalable integration model across Telegram, WhatsApp, Slack, Discord, Microsoft Teams, Google Workspace, Microsoft 365, CRMs, and task systems.

Proven Impact

• 100,000+ words embedded into long-term semantic memory.
• Responses generated in seconds.
• Personalization achieved through structured retrieval rather than repeated prompt stuffing.
• Demonstrated a practical memory-native assistant before permanent memory was widely available in mainstream LLM products.

Why It Matters

• General-purpose assistants know the world but not the individual operating them.
• Sanki demonstrates how retrieval, memory, and tools can create an assistant grounded in a real person's work.
• The project established early practical experience in vector databases, RAG, agent tooling, and memory architecture.