AethraNet

AethraNet Inspiration Intelligence has been the most consequential resource of the 21st century — and like oil before it, it has been captured by a small number of corporations. The Decentralized Human Knowledge Network proposes that intelligence should be a public utility: owned, governed, and grown by humanity collectively. The knowledge exists. It lives in millions of heads. It just is not structurally valuable to anyone collecting it, because the current AI economy concentrates value at the model and platform levels while contributors receive almost nothing.

We wanted to flip that. If a contributor's knowledge demonstrably helps answer a query, they should get paid, proportional to how useful that knowledge was, not how loud they are. Make underrepresented knowledge structurally valuable. That became AethraNet.

What It Does AethraNet is a community-owned RAG chat application built around four core mechanics:

• Contribution. Anyone can submit a document with a Stellar wallet attached. The document is embedded, chunked, and stored in a vector database.
• Retrieval. Anyone can ask a question. The system retrieves the most semantically relevant chunks, computes similarity-weighted contribution shares, and feeds them to the LLM.
• Reward. If the community has the answer, the response cites contributors and awards AETHRA points to their wallets, weighted by the similarity of each contributor's chunks to the query.
• Transparency. If the community does not have the answer, the system falls back to the model's general knowledge or a live web search, clearly labelled, with no points awarded. The topic is logged as a knowledge gap for future contributors.

Points are off-chain today and will be convertible to KNOW, an on-chain Stellar token, at network launch.

How We Built It Stack The backend is built on FastAPI, SQLite, ChromaDB, and Pydantic. LLM and embedding calls go through Google Gemini. Web fallback uses DuckDuckGo with no API key, failing closed. Wallet authentication runs on the Stellar SDK with SEP-53 challenge-response and Freighter. The frontend is React, Vite, Tailwind, and React Router. Infrastructure is Docker with docker-compose, Nginx serving the frontend, and uvicorn for the backend.

Reward Model A contributor's weight is the sum of similarity scores across all of their chunks that were retrieved. A fixed per-query point budget is then split proportionally across contributors. Reputation accrues to each contributor's weight directly, so contributors build credibility even when a point award fails.

Wallet Authentication Asking a question requires a signed Stellar wallet challenge, so every awarded point has a real destination. The auth flow is SEP-53 with server-side authoritative message construction. The server reconstructs the canonical challenge bytes from the issued nonce, and the client never sends the message back, eliminating encoding mismatch bugs.

Chat UX A familiar AI chat shell, with a collapsible sidebar of past threads, scrollable history, and a composer at the bottom. Every assistant message shows a source badge indicating whether the answer came from community knowledge, a web search, or the model's general knowledge, along with the chunks used and the exact points each contributor earned from that single message.

Challenges We Ran Into Signature Mismatch in Auth The first auth flow round-tripped the challenge through multipart/form-data, which silently normalised line endings. Signatures verified against a different byte string than the one the client signed. The fix was making the server fully authoritative: it reconstructs the canonical challenge bytes from the issued nonce, and the client never sends the message back.

Chat Scroll Bug A classic flexbox trap: a child with overflow-y set to auto inside a flex column ignores its parent's height because the default minimum height of a flex item is auto. Adding min-h-0 to the flex parents fixed it.

Out-of-Order Messages SQLite stored timestamps at second precision, so a user message and the assistant reply often shared a timestamp and were returned in arbitrary order. A rowid tiebreaker in the ORDER BY clause resolved it.

Marker Tokens Leaking Into the UI The LLM was prompted to emit ANSWER and SOURCES block markers that were not being stripped before display. Users saw raw markers in their chat. The fix was returning two fields from the RAG pipeline: a clean answer text for the message bubble and a richer formatted response for debug or export.

Routing Greetings Through RAG A naive RAG pipeline routes every greeting into vector search and gets strange results. A small-talk classifier with its own reply path handles conversational openers with no embedding call, no LLM call, no reward, and no gap logged.

On-Chain Reward Latency On-chain payouts per query produced 3 to 8 second tail latency, transient failures that made contributors question whether they had been paid, and testnet rate limits that broke demos. The solution was off-chain AETHRA points with a stub claim endpoint returning HTTP 501 and a clear token-launch-coming-soon message. Same economic model, decoupled from chain finality, ready to switch on at launch.

Accomplishments We Are Proud Of • A real end-to-end loop. Contribute, embed, retrieve, answer, reward: every step works, with the same Stellar identity tying it together. • Three honest answer modes. Community, Web, and General, each labelled, each with its own reward semantics. The app never pretends it knows something it does not. • A wallet auth flow that does not betray users. SEP-53 with server-authoritative challenge bytes eliminates silent encoding bugs. • A reward model that scales beyond the demo. Off-chain points with a planned on-chain claim means onboarding thousands of contributors without paying for thousands of transactions. • A clean, familiar chat UI built around RAG-first transparency: provenance shown by default, not buried in a developer tool. • A working knowledge-gap registry. Unanswered questions become first-class objects the community can target.

What We Learned • Trust beats throughput. Hiding the source made answers feel magical and untrustworthy at the same time. Surfacing source badges raised satisfaction on the same answers. • Off-chain points are a feature, not a compromise. Decoupling reward UX from chain finality kept the economic model without the latency tax. • RAG without graceful fallback is hostile. A pure community-RAG system refuses greetings, refuses arithmetic, and shrugs at common knowledge. Small-talk, general, and web fallbacks, each labelled, fixed it. • Cryptographic UX is a content-policy problem. A signature failure caused by a normalised newline is technically a user error and emotionally a betrayal. Most of the auth work was about making the human experience of Ed25519 succeed. • CSS bugs live two layers up the tree. min-h-0 is the answer more often than it should be.

What Is Next for AethraNet • Launch KNOW on Stellar mainnet and wire the claim endpoint to burn points and mint or transfer tokens. • Open a knowledge-gap bounty board: anyone can fund a gap, and the contributor who fills it splits the bounty with retrievers whose work the LLM still cites. • Multilingual ingestion and retrieval, prioritising Luganda, Swahili, Hausa, Yoruba, and Amharic. • A provenance receipt per answer: a downloadable, signed manifest of which chunks, which contributors, and which similarity scores produced it. • Mobile-first contribution flows using voice to transcript to chunk, for contributors whose knowledge lives in conversation rather than documents.

The bet behind AethraNet is simple: an internet of AI built only from corporate scrapes will keep producing models that do not see most of humanity. The fix is not more scraping. It is giving the people who hold the knowledge a reason to share it and a share of what it builds.

Built With

• and
• and-pydantic.-llm-and-embedding-calls-go-through-google-gemini.-web-fallback-uses-duckduckgo-with-no-api-key
• and-react-router.-infrastructure-is-docker-with-docker-compose
• chromadb
• failing-closed.-wallet-authentication-runs-on-the-stellar-sdk-with-sep-53-challenge-response-and-freighter.-the-frontend-is-react
• for
• nginx-serving-the-frontend
• sqlite
• tailwind
• the
• the-backend-is-built-on-fastapi
• uvicorn
• vite