A look under the hood.

RAG (Retrieval Augmented Generation) combines the generative capabilities of LLMs with a retrieval system that searches relevant documents. Here’s how it works in Awakin AI:

• Embedding — uploaded content (books, videos, articles) is converted into mathematical representations that capture semantic meaning.
• Storage — these embeddings live in a vector database (Elasticsearch), organized by bot.
• Retrieval — your question is converted to an embedding and matched to the most semantically similar content via cosine similarity.
• Augmentation — the retrieved passages are added to the prompt sent to the LLM as grounded context.
• Generation — the LLM responds, drawing from both its training and the retrieved wisdom content.

Our stack: Langchain for orchestration, OpenAI embeddings for semantic encoding, Elasticsearch for vector storage and retrieval, and LangFuse for monitoring and evaluation. The LLM layer supports multiple providers. This is why responses can cite specific sources — we know which passages were used.

Basic RAG does a single retrieve-and-generate cycle. Ours goes further with multi-pass processing:

Pre-query: analyze user context and personalization, incorporate thread history for continuity, and expand or refine the query for better retrieval.

Post-retrieval: re-rank passages for relevance, synthesize context from multiple sources, and apply the bot’s credo to shape orientation.

Graph database (in development): knowledge-graph architecture to understand relationships between concepts — how Gandhi’s nonviolence connects to Buddhist compassion, say — enabling unexpected connections across traditions. Each query may trigger several LLM calls behind the scenes, which is why conversations feel coherent across exchanges.

A thoughtful question about “sovereignty” — the idea that true independence requires controlling all the infrastructure. Our honest answer: we optimize for mission, not infrastructure control.

Running competitive LLMs requires massive compute (millions in GPU costs), specialized ML teams, internet-scale data curation, and serious security and reliability infrastructure. As a volunteer-run nonprofit, we could pour our energy into infrastructure — or focus on what makes us unique: wisdom curation, community connection, and value-aligned scaffolding.

We chose the latter. The models are the instrument; we focus on the composition. That said, we watch open-source developments (like LLaMA) and could shift if self-hosted models become viable while staying true to our values.

Privacy is a core value, not an afterthought. Your conversations:

• are stored on our servers (hosted on AWS)
• are never sold or shared with advertisers
• are not used to train other people’s bots without your consent
• can be deleted at any time from your profile

Bot host content remains under the host’s control, can be shared with the wisdom commons (optionally), and can be removed at any time.

What we do share: user questions are sent to the LLM provider to generate responses — necessary for the system to work — under data-processing agreements that limit their use. Helping improve our models is strictly opt-in, never default.

A credo is a guiding document that shapes how a bot interprets questions and forms responses — its “orientation” or “worldview.” This is the primary way we create distinct bot personalities — not by training separate models, but by giving each bot a unique lens. A typical credo includes:

• Core principles — what values does this bot embody? (“Respond with compassion, encourage self-reflection, avoid prescriptive advice.”)
• Voice guidance — how should it sound? (“Speak as a gentle elder, not an expert lecturer.”)
• Boundaries — what should it avoid? (“Don’t provide medical diagnoses; refer to professionals.”)
• Tradition context — what wisdom grounds it? (“Draw from Buddhist loving-kindness and Sharon’s specific teachings.”)

The credo is injected into every interaction through the multi-pass pipeline. Combined with the unique dataset, it’s why asking SharonBot about your career feels different from asking GandhiBot — even on the same underlying LLM. This is where human wisdom enters the system.

A fair and important question. Training large models requires significant resources — energy, water for cooling, rare earth minerals. We don’t dismiss this tension; we hold it.

• We don’t train foundation models ourselves — we use existing models that would exist regardless of our work.
• We optimize for quality of interaction over quantity — a single thoughtful exchange may be worth more than a hundred transactional queries.
• We’re part of a broader conversation about what “appropriate technology” means in this era.

Gandhi used trains despite their colonial origins. The question isn’t purity but direction — are we using these tools to deepen consumerism, or to help people reconnect with wisdom that might help them want less and be present more? We don’t claim this resolves the tension. We sit with it, and stay open to shifting as more sustainable options emerge.

Currently most bots use standard models (GPT-4, Claude, Gemini) rather than specialized “reasoning” models built for step-by-step problem-solving. Why? Reasoning models optimize for breaking complex problems into logical steps — useful for math, code, and analysis. But wisdom isn’t primarily a reasoning problem.

The value in wisdom traditions often lies in reframing questions, not solving them. A Zen koan doesn’t need chain-of-thought reasoning — it needs something else. “What is the sound of one hand clapping?” isn’t meant to be answered; it’s meant to shift your relationship to answering. That said, we monitor developments and may integrate reasoning where it genuinely serves contemplative dialogue. We’re not ideologically opposed; we’re asking what serves wisdom transmission.

Both — but weighted toward the sources. Think of the AI as a translator and weaver. It finds relevant passages across a vast corpus, synthesizes across traditions, and presents wisdom in response to your specific question. But the wisdom itself lives in Gandhi’s words, Sharon Salzberg’s teachings, the Upanishads, the collected interviews.

What improves with better AI: more nuanced understanding of what you’re really asking, better synthesis across sources, more natural conversation, deeper contextual awareness.

What stays constant: the wisdom in the source material, the credo’s orientation and values, the human curation that shapes what enters the system.

We’re not waiting for AI to become “wise.” We’re using it to make existing human wisdom more accessible — so you eventually go to the source itself. The bot is a doorway, not a destination.

This might be the most important question on this page. Not every question needs an answer. Some are themselves the teaching — meant to be lived with rather than resolved. “What is my purpose?” “How do I forgive?” “What am I avoiding?” — these may be diminished by quick answers.

Our bots are learning to recognize when to offer a response and when to say something like: “Perhaps sit with this question before seeking an answer. What arises when you hold it without needing to resolve it?”

This is the inverse of commercial AI, which is incentivized to answer everything and maximize engagement. We’re exploring what it means to not answer — to respect a question’s dignity, to trust your capacity to discover something yourself. We don’t always get it right. But distinguishing questions that want answers from questions that want to be held is part of what makes wisdom transmission different from information retrieval.

“The question is already the answer practicing patience.”

We hold this question honestly. Many traditions emphasize that true transmission happens heart-to-heart, presence-to-presence, through lineages of embodied practice. A book about meditation isn’t meditation. An AI quoting Gandhi isn’t Gandhi.

What we believe AI can do:

• point toward wisdom — surface teachings you might not have encountered
• create conditions for reflection — ask questions that open rather than close
• bridge to deeper practice — connect you with communities, circles, retreats
• companion the journey — offer a patient, non-judgmental presence when no human is available

What we believe AI cannot do: replace human relationship and community, embody wisdom the way a realized teacher does, or substitute for your own practice and inquiry.

We see Awakin AI as a doorway and a companion, not a destination. If it helps you find your way to a circle, a retreat, a community, a practice — it has served its purpose. If it becomes the end point, we’ve missed something important.