laterbase

A lightweight, self-hosted read-later link collector. Share a link from WhatsApp (via a Hermes agent), paste it into the dashboard, and laterbase fetches the page, categorizes it with an LLM, generates a summary, and files it into a group — all in a clean, minimal dashboard.

WhatsApp → Hermes → POST /api/links → River job queue → fetch → LLM categorize → store
                                                   ↓
                                           Dashboard (React SPA)

Features

• Capture anywhere — send link <url> to a Hermes WhatsApp agent, or add links via the dashboard's Add Link dialog
• Automatic processing — each link is fetched, cleaned, categorized (article, github, youtube, docs, tools, ...), summarized (gpt-4o-mini), and assigned to a group by a River background job
• Honest statuses — inaccessible, paywalled, or failed pages are clearly labeled; content is never fabricated
• Groups — organize links into color-coded groups (articles, github, docs, videos, tools, uncategorized) or create your own
• Content preview — read saved page text inline, with an "Open in browser" fallback for sites that block embedding
• Notes — attach free-form notes to any link
• Ask your archive — press ⌘K (or the Ask button) to ask a natural-language question; gpt-5.6-luna ranks your saved links by relevance, streams a cited answer, and lists the matching links — every claim traceable to something you actually saved
• Pluggable fetcher — simple HTTP + readability extraction by default, with an optional Crawl4AI service for JS-heavy pages

Tech Stack

• Backend: Go (chi router, pgx pool, goose migrations, River job queue)
• Frontend: React + Vite, Tailwind v4, TanStack Query, React Router
• Database: PostgreSQL 17
• LLM: gpt-4o-mini (categorization/summarization) + gpt-5.6-luna (Ask feature) via OpenAI API
• Optional: Crawl4AI (Docker profile) for JavaScript-heavy sites

Quick Start

Requires Docker (Docker Compose v2).

cp .env.example .env   # then fill in LLM_API_KEY
docker compose up -d   # starts postgres + api (air hot-reload) + frontend (vite dev)

Optionally start Crawl4AI:

docker compose --profile crawl4ai up -d

Service	URL
Dashboard	http://localhost:5173
API	http://localhost:8080
PostgreSQL	localhost:5433
Crawl4AI (optional)	http://localhost:11235

Database migrations run automatically on API startup via goose. The backend hot-reloads on .go changes (Air) and the frontend uses Vite HMR.

Configuration

Environment variables (see .env.example):

Variable	Required	Description
LLM_API_KEY	Yes	OpenAI API key for categorization, summarization, and Ask
LLM_MODEL	No	Model for link processing (default: gpt-4o-mini)
ASK_LLM_MODEL	No	Model for the Ask feature (default: gpt-5.6-luna)
CRAWL4AI_URL	No	Crawl4AI endpoint; empty = simple HTTP fetcher

Note: newer OpenAI models (like gpt-5.6-luna) only support max_completion_tokens and the default temperature — the Ask pipeline already accounts for this.

The fetcher mode can also be toggled at runtime:

curl -X PUT localhost:8080/api/config \
  -H 'Content-Type: application/json' \
  -d '{"fetcher_mode": "crawl4ai"}'

Hermes Integration

Add the following to your Hermes agent's SOUL.md to capture links from WhatsApp:

### `link <url>` or `save <url>`
- POST the URL to the laterbase backend at http://localhost:8080/api/links
- Body: {"url": "<the url>"}
- Acknowledge with: "Saved. I'll fetch and summarize it shortly."
- If the API is unreachable, respond: "laterbase isn't running. Start it first."

API

All routes are under /api (plus GET /health). Highlights:

Method	Path	Description
POST	/api/links	Ingest a link — body: {"url": "..."}
GET	/api/links	List links — supports ?group_id=, ?status=, ?search=
GET	/api/links/:id	Single link with full extracted text
PUT	/api/links/:id	Update group, title, source type, or notes
DELETE	/api/links/:id	Delete a link
POST	/api/links/:id/reprocess	Re-run fetch + categorize + summarize
GET / POST / PUT / DELETE	/api/groups[...]	Manage groups
GET	/api/stats	Counts by status, group, and source type
GET / PUT	/api/config	Read/update fetcher mode
POST	/api/ask	Ask a question over the archive — body: {"question": "...", "include_full_text": false}; responds with SSE (citations → delta → done)
GET	/health	DB ping + queue status

See cmd/server/main.go for the full route table.

Project Structure

laterbase/
├── docker-compose.yml
├── .env.example
├── PLAN.md                  # Original build plan & design decisions
├── backend/
│   ├── cmd/server/main.go   # Route table & wiring
│   └── internal/
│       ├── config/          # Env config
│       ├── db/              # pgx pool, queries, embedded goose migrations
│       ├── handlers/        # HTTP handlers (chi)
│       ├── fetcher/         # Simple HTTP + Crawl4AI implementations
│       ├── categorizer/     # gpt-4o-mini call
│       ├── ask/             # Ask pipeline: rank links + stream grounded answer
│       ├── workers/         # River ProcessLink job
│       └── content/         # HTML sanitizer
└── frontend/
    └── src/
        ├── pages/           # Dashboard, LinkDetail, Groups
        ├── components/      # LinkCard, ContentPreview, etc.
        ├── hooks/           # TanStack Query hooks
        └── api/             # Axios client + SSE ask client

Documentation

• PLAN.md — full architecture, data model, and design rationale
• PRODUCT.md — product context and positioning
• DESIGN.md — visual system and tokens