Build

Architecture

One server, three partners doing real work, and a fallback for every agent role.

Konkyo's system: the web app and worker on one server, with Postgres and pgvector, publish every file to Zetaris, use three Meterless engines, and call NVIDIA Nemotron and OpenAI with a fallback and a reserve for every agent role.
The whole system: what runs where, and what each partner does.

The pieces

PieceWhat it doesIn the code
Web appThe app, its API (/api/v1/ask), and the answers, streamed as the agents workapps/web
WorkerReads every file, runs deep searches, makes study sets, runs agents and audio briefsapps/worker
Postgres + pgvectorPassages with their Nemotron vectors and full-text index, chats, runs, study cardspackages/db
Object storageEvery file and its PDF rendition, so one viewer can highlight any citationpackages/storage
ZetarisThe evidence store: every passage, Zetaris's own reading of every PDF, and the SQL that checks every quotepackages/zetaris
MeterlessH-MEM (your corrections), World Model (which source says what), Markovian (screens deep-search finds in bounded steps)packages/meterless, packages/agents
ModelsOne place for every model call: a deadline and a rate limit for each, and for every agent role a fallback on the other provider and a reservepackages/models

One answer

One answer: the Planner reads H-MEM and plans searches, the Retriever finds passages, the Writer drafts one finding per sentence, the Verifier checks every quote in Zetaris in one SQL statement and sends unsupported sentences back, the Analyst records agreement in the World Model, and only what passed is answered, with its receipt.

Step by step, with the models: How an answer is checked.

Where a file goes

Where a file goes: read by kind (a PDF directly, a scan with OCR, Word, slides, web pages and email through Gotenberg to PDF, recordings and caption-less YouTube with ElevenLabs Scribe), split into passages with Nemotron Embed vectors, full text and a Lightning summary, published to Zetaris, and Ready only when Zetaris's counts match.
  1. Upload or import: a PDF, Word file, slides, web page, YouTube link, recording, email, or a Gmail, Drive or Notion item.
  2. Readable text: a PDF is read as it is, and its scanned pages get OCR; Word, slides, Markdown, web pages and email become a PDF through Gotenberg; recordings (and YouTube without captions) are transcribed with ElevenLabs Scribe, cited to the second.
  3. Passages: split with their page and position, embedded with Nemotron 3.5 Embed, summarised with Nemotron 3.5 Lightning.
  4. Into Zetaris: the passages and the PDF are published; Zetaris reads the PDF itself.
  5. Ready: only once Zetaris answers for it. From then on, every quote from it is checked there.

Where it runs

One Coolify server runs the web app, the worker, Postgres, object storage, Gotenberg and this docs site (a static export), with Zetaris beside them. konkyo.xyz and docs.konkyo.xyz are what's described here. To run your own: Self-hosting.

On this page