Post 11 of the Pepper & Carrot AI flipbook series — the alternative deploy. Post 10 put the reading companion on five clouds with a Modal GPU serving Ollama, because the series is about local-first inference. This post takes the same app and ships it without a GPU at all: chat on the Anthropic Messages API, embeddings on Voyage AI. The whole point is that it's a *configuration* change, not a code change — the provider abstraction from Post 3 was built for exactly this, and the only new code in the repo is documentation. The interesting parts are the trade the swap makes (cost and latency for the local-first thesis) and the one real gotcha nobody warns you about: a managed embeddings model lives in a different vector space, so the search index has to be rebuilt before the first deploy or retrieval silently returns garbage.

Post 10 of the Pepper & Carrot AI flipbook series — the deploy. The flipbook, the spoiler-safe RAG, the world graph all run beautifully on the developer laptop the first nine posts built around. This one puts the same architecture on the public internet for roughly the price of a coffee a month — Cloudflare Pages for the static frontend, Fly.io for the FastAPI backend, Modal for the GPU-served Ollama, Neon for managed Postgres, Cloudflare R2 for the image bytes. The provider abstractions from Post 3 finally cash in: the backend doesn't notice that Ollama moved off localhost, the storage swap is one env var, the database URL is one secret. The new code is small (a boto3-backed R2Storage finally lands behind the Post 3 Protocol, a Dockerfile, a fly.toml, three short infra scripts) — the harder work is the architectural judgement about which seams to draw and where to put the cold start the budget can afford.

Post 9 of the Pepper & Carrot AI flipbook series. The chat layer answers questions about pages and the wiki; now we add a third affordance — a spoiler-aware knowledge graph of the comic's world, rendered as an in-reader overlay — plus a third Claude Code skill that closes the loop back to the chat. A second skill walks the wiki sources + the per-page description JSONs and writes a durable YAML pair. A FastAPI route filters the graph with a Postgres row-value comparison so an edge whose own debut is past the reader's cursor doesn't leak even when both of its endpoints are visible. Two response modes — focus (on-page characters + 1-hop structural neighbors) and full (the whole spoiler- safe world) — share the same boundary. A React + xyflow overlay renders circular avatar nodes with kind-based SVG fallbacks, a kind-filter bar, kind-colored edges that brighten on the selected node, soft fade-in for newly-revealed entities, and an "Ask in wiki mode" click that round-trips back through the chat panel. A third skill — summarize-wiki — authors one tight ~150-word .md per entity so that "Ask in wiki mode" for a minor character like Truffel or a coven like Magmah actually works against qwen2.5:7b, instead of the 30 KB multi-entity articles blowing past the prompt-hardening guarantees from Post 8.

Post 8 of the Pepper & Carrot AI flipbook series. Post 7 left a streaming chat panel and an honest admission: the engineering guarantees structure and safety, but it doesn't guarantee taste. This post is the prompt-engineering pass that closes that gap on a 7B local model — a markdown stripper on every piece of text entering the prompt, a closed-world grounding contract, a page-mode anti-recitation block, a strict response-format cap, a much sharper suggestion-chip prompt with bad/good examples, and react-markdown in the chat panel as the belt-and-suspenders safety net.

Post 7 of the Pepper & Carrot AI flipbook series. Post 6 left a spoiler-safe chat pipeline you could only reach with curl. Now we put it in the browser: tokens stream over Server-Sent Events into a React chat panel, the user picks page or wiki mode per message, and two follow-up suggestion chips render below each answer — generated by a second model call, constrained to a JSON schema, and validated server-side before a single chip reaches the DOM. Plus a light wiki ingestion path so wiki mode has something to say.

Post 6 of the Pepper & Carrot AI flipbook series. The flipbook from Post 5 knows which page you're on. Now we build the chat pipeline that answers questions about that page — and we make spoiler safety a property of the database query, not a line in the prompt. Build a RetrievalService whose Chroma filter is derived from server-side reading progress, wire it into a FastAPI chat endpoint, drive it with curl, and prove the boundary holds even when the user tries to jailbreak it. No chat UI yet — that's Post 7.

Post 5 of the Pepper & Carrot AI flipbook series. With one episode sitting in Postgres + LocalStorage from Post 4, it's time to surface it. Build two typed FastAPI routes that resolve relative storage keys into absolute URLs at response time, and wire up a real page-flipping flipbook with React + StPageFlip — single page in portrait, two-page spread in landscape. By the end you have an episode picker plus a flipbook rendering real data from your local backend.

Post 4 of the Pepper & Carrot AI flipbook series. The comic is images, not text — so before any RAG can happen, every page needs a description. This post walks through using a Claude Code skill as the vision provider for the ingestion pipeline of this portfolio-project specifically: no per-call API cost beyond the Claude Code subscription, auditable JSON artifacts on disk, same Claude model as Anthropic's hosted vision API. By the end, one full episode is ingested into Postgres + ChromaDB + local storage. The right vision provider is context-specific — local VLM, hosted API, and Claude Code each win under different constraints (budget, whether the pipeline runs unattended, throughput) — and the post includes a decision matrix mapping each constraint to the right choice.

Post 3 of the Pepper & Carrot AI flipbook series. Build three typed Protocol interfaces — Storage, EmbeddingClient, ChatClient — and the factory that picks the right implementation from a .env file. By the end you have LocalStorage serving images end-to-end and a working embedding client producing real 1024-dim vectors against the local Ollama you set up in Post 2.

Post 2 of the Pepper & Carrot AI flipbook series. Stand up the local workshop: Postgres in Docker, Ollama serving qwen2.5:7b and bge-m3, a typed FastAPI scaffold that passes mypy strict, the first Alembic migration, and one downloaded episode on disk. By the end you have the workshop every later post is built on.