What you can build
Concrete setups that K-AI is designed for.
A private chat service
Run Open WebUI (or any OpenAI-compatible front end) against one K-AI endpoint. Every chat profile appears as a model; people pick one and K-AI loads it, queueing requests instead of failing.
- Local GPUs first, rented pods when you choose
- One key per front end, revocable any time
Large models without a standing bill
Mark a profile as allowed on RunPod, set a monthly cap and an idle-stop time. Pods start when you need them, load from a network volume, and delete themselves when idle.
- Spend latch at the cap
- Reload a profile in about two minutes, no new pod
A grounded assistant over your documents
Upload PDFs, ebooks, Word and Excel files into a knowledge base and bind it to a profile. Answers draw on the most relevant passages and the relationships between entities, with a reason for each link.
- One model profile per knowledge base, sharing one loaded model
- Inject, tool or both
Reference libraries
Turn a shelf of technical books, rulebooks or manuals into searchable, graph-linked knowledge. Keep separate bases for separate subjects and give each its own model entry in your chat app.
- Duplicate detection by hash
- OCR for scanned pages
Coding agents with repository memory
Add a git repository as a source and connect an agent over MCP. It can search code and docs, find a symbol, and walk what calls or imports it. Webhooks keep the index current.
- Tree-sitter parsing for many languages
- Infrastructure files (Dockerfile, Compose, Terraform) understood
Scoped keys for people and tools
Give a script a models-only key, a teammate read access to one knowledge base, and an automation read-write on another. See who used what, with token counts, for 90 days, without storing prompts.
- Expiry and instant revoke
- CSV export
A spare gaming PC as the knowledge node
Run embedding and extraction on a 16 GB Windows GPU through Docker Desktop while chat runs on rented pods. Everything stays reachable through one router.
- One-shot bootstrap script
- Embed and extract share the GPU
Evaluating models
Search Hugging Face from the panel, see whether a model fits your GPUs and at what context length, download it, and create a profile pre-filled with sensible settings.
- Fit verdicts: fits, tight, too big
- Tool and reasoning parsers by family
Overnight quality upgrades
Ingest quickly with a small extractor by day; let idle knowledge nodes re-extract with a bigger model at night so the graph keeps improving.
- Configurable window and idle period
- Replaces a chunk's graph only on success