ai-engineer-wiki

name

ai-engineer-wiki

description

Field-tested knowledgebase distilled from AI Engineer conference talks for building AI software and building with AI — agents, coding agents, RAG/retrieval, evals, inference, context engineering, MCP/tools, models, voice, vision, security, AI product strategy. Use when answering a question or making a design decision about building AI/LLM systems and you want grounded patterns, tradeoffs, and failure modes with source citations; also use when another agent needs grounded AI-engineering knowledge.

AI Engineer Wiki

A curated, source-backed wiki distilled from AI Engineer YouTube talks. Use it to ground AI-engineering questions and design decisions in field-tested patterns instead of answering from memory alone. It is read-only — you consume it over the network; you never edit it from here.

It is fetched on demand from GitHub raw URLs — no clone, no local copy. The source of truth is the main branch of tschuehly/ai-engineer.

BASE=https://raw.githubusercontent.com/tschuehly/ai-engineer/main/data/ai_engineer_youtube/wiki

Four page types live under $BASE:

• topics/<slug>.md — 21 synthesized articles. Broad/orienting questions ("how should I think about evals?").
• concepts/<slug>.md — ~780 atomic, source-backed pages. Specific claims, patterns, tradeoffs, failure modes. Each cites the originating video with timestamps.
• sources/<file>.md — one provenance note per video. Use when the user wants the original talk, speaker, or quote.
• indexes/concept-index.md — every concept title grouped by topic, with its link. The searchable map.

Topic routing table

Map the question to one or more topics first — no fetch needed for this step. Slugs are stable.

slug	covers
agents	agent architecture, autonomy, tool use, multi-agent orchestration
coding-agents	coding agents, harnesses, codegen, review, agent-legible codebases
workflows	workflows vs agents, structured pipelines, orchestration
context-engineering	context windows, memory, prompt/context management
tools	tool & MCP design, tool catalogs, agent connectivity
retrieval	RAG, search, embeddings, knowledge bases
evaluation	evals, benchmarks, agent-trajectory evaluation, LLM judges
models	model families, training, capabilities, post-training/RL
inference	serving, latency, throughput, caching, quantization, cost
edge-inference	on-device / edge model inference
infrastructure	platforms, sandboxes, agent runtimes, deployment infra
security	agent/LLM security, permissions, MCP risks, sandboxing
voice-agents	voice/speech agents, realtime audio
vision-ai	vision models, multimodal perception, OCR
generative-media	image/video/audio generation, diffusion
robotics	embodied agents, VLA models, robot control
product-strategy	AI product strategy, taste, adoption, org design
ai-monetization	pricing, business models, ROI for AI products
architecture-copilots	copilots, AI architects, design assistants
business-intelligence	AI for BI, analytics, data querying
healthcare-operations	AI in healthcare / clinical operations

Retrieval procedure

Use curl -fsSL (exact, greppable bytes). Fetch only what you need; follow links to go deeper.

1. Route. Pick the relevant topic slug(s) from the table above.

2. Search the index for candidate concept pages. Grep the concept-index for your keywords (run several greps with synonyms for recall — the index is large, so let grep filter it, don't pull it whole into context):

   curl -fsSL "$BASE/indexes/concept-index.md" | grep -iE 'eval|benchmark|llm.?judge'
   

   Each hit is a - [Title](../concepts/<slug>.md) line — the <slug>.md is the exact path you need.

3. Read. For orientation, fetch the topic article; for specifics, fetch the concept pages whose titles matched:

   curl -fsSL "$BASE/topics/evaluation.md"
   curl -fsSL "$BASE/concepts/evaluate-agent-loops-with-correctness-cost-latency-and-step-counts.md"
   

4. Follow links. Concept pages end with Related concepts, Related topics, and Sources. Fetch a few of these when the first pages don't fully answer the question. Stop when you have enough — you do not need to read the whole wiki.

5. Answer with citations. Synthesize across the pages you read. For every claim, name the concept page and the video provenance it cites (speaker, talk title, timestamp) so the answer is verifiable. When advising a coding agent on a build decision, give the concrete pattern/tradeoff and link the concept page it came from.

Rules

• Never fabricate a concept path. Slugs are not a clean transform of titles (e.g. "Target High-Value AI Verticals as Capability Matures" → target-high-value-ai-verticals.md). Always copy the exact <slug>.md from a link in the index or a topic page. If you didn't see the link, search the index first.
• Resolve relative links to raw URLs: both ../concepts/x.md and concepts/x.md become $BASE/concepts/x.md; ../sources/y.md becomes $BASE/sources/y.md.
• Prefer concept pages for specific claims (they are atomic and carry source timestamps) and topic pages for synthesis/orientation.
• Read-only. Do not attempt to write back into the wiki from a consuming project. To add knowledge, work in the tschuehly/ai-engineer repo itself.
• WebFetch is the fallback only where no shell/curl is available. Prefer curl so the index stays greppable and content isn't summarized away.
• On 404 / empty results: the wiki may not be merged into main yet, or the path is wrong. Re-derive the path from a fresh index grep; if $BASE/index.md itself 404s, tell the user the wiki likely isn't on main yet.