TOJI TOJI
ARCHITECTURE · cogito

cogito architecture

cogito is an LLM-based ingest pipeline that classifies any file — a PDF, a Markdown note, an unstructured scrap — into an entity whoami can use. It's a separate gRPC service so it can remove whoami's rigid directory-structure requirement, and it's the shared foundation behind report pipelines like jaso (resume), bizplan, and company — including the resume PDF content referenced elsewhere on this site.

Project overview →

Why

When generating responses with an LLM, output quality turned out to be lower and less consistent than expected. To fix both the quality and consistency problems, this project integrates a question pipeline, a verification pipeline, and a composition pipeline together, aiming to improve reliability through more consistent, higher-quality responses.

01

How the internals are decomposed

cogito-structure.yaml

whoami enters over gRPC (cogito_service · prechunk_sink); the worker (runner · sink_coordinator · sweeper) hands off asynchronously to pipelines (taxonomy · jaso · bizplan · company), which call agents for LLM invocations. A Postgres-backed registry tracks processing state.

cogito component structure: whoami calls cogito_service and prechunk_sink over gRPC, which hand off to worker, then pipelines, then agents for LLM calls, writing to a Postgres registry and reaching gopedia

cogito / structure.yaml

02

How it deploys, and what it depends on

cogito-infra.yaml

Scale-to-zero by default — the ArgoCD Image Updater auto-scales it when it detects a new registry digest. LLM inference calls Ollama on a bare-metal GPU node; knowledge search calls gopedia-svc; entity storage calls whoami-api.

cogito infra: toji-cd ArgoCD Image Updater watches artifacts.toji.homes for a new digest and syncs the cogito-svc Deployment, which calls Ollama, gopedia-svc, whoami-api, and otel-collector

cogito / infra.yaml

03

What path a single file takes

cogito-what.yaml

Text is extracted from the uploaded file, an LLM classifies it into a selector, then frontmatter is generated/merged. Content-hash based, so re-uploads upsert into whoami without creating duplicates.

cogito ingest pipeline: file upload flows through text extraction, LLM classification, frontmatter generation/merge, then handoff to whoami

cogito / what.yaml

04

How a commit becomes a deployment

cogito-ops.yaml

A git push webhook triggers toji-ci (Tekton), which delegates the build to dagger-engine and pushes the image to the registry. The ArgoCD Image Updater polls the registry directly for the digest change — unlike goquest/lymphhub's manifest-writeback flow, no PR is needed for the deploy to land.

cogito CI/CD: git push triggers toji-ci Tekton, which delegates to dagger-engine, pushes an image to the registry, and ArgoCD Image Updater detects the digest change and syncs the cogito-svc Application

cogito / ops.yaml

rendered via archview · d2lang/d2 · ELK layout