🎓
IBM/Coursera coverage
IBM "RAG and Agentic AI" (Coursera) Coverage ↔ This Course
Detailed mapping of the IBM professional program on Coursera (10 courses) against this repository's own curriculum (M0–M11). The IBM syllabus was coverage reference only — we did not follow that course or replicate its labs. We built our own material, in Spanish, with a tri-modal approach (design + Python from scratch + real frameworks) and go beyond it in production, design, and industry cases. See
PLAN.md §11andHANDOFF.md §2.
Index
- Context and scope
- Main table: 10 IBM courses → modules
- IBM skills → where they are acquired
- Extras this course adds (and the IBM program does not emphasize)
- Checklist: if you come from the IBM program, start with…
- Conclusion
1. Context and scope
The IBM "RAG and Agentic AI" program on Coursera consists of 10 courses that progressively cover GenAI applications with LangChain, RAG, vector databases, advanced RAG, multimodal, agents, agentic frameworks (LangGraph, CrewAI, AutoGen, BeeAI), MCP, and an integrative capstone.
What we did with that reference:
| Aspect | IBM program | This course |
|---|---|---|
| Role of IBM syllabus | Course to follow | Coverage checklist — ensure no topic is missing |
| Material | Coursera labs | Our own guides, exercises, workshops, and solutions in rag-training/ |
| Code | Mostly frameworks | Tri-modal: ① design → ② scratch (stdlib) → ③ real framework |
| Use cases | Generic / course datasets | 10 RAGorbit industry templates (airline, banking, healthcare…) |
| Production | Minimal | Full M9: guardrails, HITL, Kafka, Temporal, AI Security |
| Final evaluation | Coursera project | M11 Capstone: rebuild 3 templates + design architecture + exam |
RAGorbit node coverage: all 13 categories (model, loader, ingest, store, retrieval, query, logic, agent, tool, guardrail, hitl, observability, io) are mapped to modules in PLAN.md §7. This document cross-references that coverage with the IBM syllabus.
2. Main table: 10 IBM courses → modules
Each row breaks down IBM course topics, where they are covered in this course (module + folder), and notes on depth or differences.
Course 1 · Develop GenAI Apps
IBM topics: basic LangChain, prompt patterns, JSON output, model evaluation, deployment with Flask.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| LLM fundamentals (tokens, context, temperature) | M1 | ../01-fundamentos/ |
+ embedding intuition; model choice latency/cost/quality |
| Prompt patterns (system/user, few-shot, CoT, in-context learning) | M1 | ../01-fundamentos/ |
Prompt Patterns skill; Claude/OpenAI/Gemini/Llama comparison |
| Why RAG vs fine-tune vs prompting | M1 | ../01-fundamentos/ |
Minimal RAG pattern ~40 lines (scratch) + LangChain (③) |
| LangChain / LCEL (basic chains) | M1, M5 | M1 + ../05-generacion-y-logic/ |
Depth in M5: query engines, structured output |
| JSON output / structured output | M5 | ../05-generacion-y-logic/ |
JSON Schema, requireCitations, instructor/outlines |
| Evaluation and model selection | M1, M5 | M1 + M5 | RAGAS, TruLens, DeepEval, promptfoo in M5 |
| OpenAI APIs (and alternatives) | M1 | ../01-fundamentos/ |
OpenAI API skill; provider:model format without lock-in |
| Flask UI / lightweight deployment | M9 | ../09-produccion-y-seguridad/ |
Flask + Gradio/Streamlit; FastAPI/SSE targets |
Course 2 · Build RAG Apps
IBM topics: end-to-end RAG pipeline, Gradio, LlamaIndex vs LangChain comparison.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| RAG pattern (retrieve → synthesize) | M1, M5 | M1 + M5 | Template 09-RRHH |
| Loaders and document preparation | M2 | ../02-ingesta/ |
PDF, tabular, web, SQL, S3; chunking and metadata |
| Embeddings and indexing | M3 | ../03-embeddings-y-stores/ |
Chroma, FAISS, pgvector; see course 3 |
| Retrieval and generation with context | M4, M5 | M4 + M5 | Hybrid, rerank, mandatory citations |
| LlamaIndex (readers, query engines) | M2, M4 | M2 + M4 | Hands-on: chunk legal PDF + LlamaIndex retriever |
| LangChain vs LlamaIndex (trade-offs) | M2, M4, M5 | M2–M5 | Table in tecnologias-comparadas.md |
| Gradio to prototype RAG | M9 | ../09-produccion-y-seguridad/ |
Conversational UI; Streamlit too |
Course 3 · Vector DBs for RAG
IBM topics: ChromaDB (CRUD operations), vector similarity, recommendation systems.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| Embeddings (dimensions, normalization, metrics) | M3 | ../03-embeddings-y-stores/ |
Cosine, dot, L2; OpenAI vs Cohere vs local BGE/E5 |
| Vector indexes (HNSW, IVF, flat) | M3 | M3 | HNSW vs flat comparison in workshop |
| ChromaDB: add/update/delete/manage | M3 | M3 | Vector DBs skill; store.chroma node |
| FAISS by hand | M3 | M3 | Scratch implementation + comparison with Chroma |
| Vector store vs traditional DB | M3 | M3 | pgvector, Qdrant, Pinecone, Weaviate, Milvus |
| Recommendation systems with embeddings | M3 | M3 | Explicit IBM extra; item↔item and user↔item analogy |
| Persistence and collections | M3 | M3 | Template 09 (Chroma), 02 (pgvector) |
Course 4 · Advanced RAG
IBM topics: FAISS, HNSW, advanced retrievers, Gradio.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| FAISS / HNSW in depth | M3, M4 | M3 + ../04-retrieval-y-query/ |
Indexes in M3; retrievers that consume them in M4 |
| BM25 / keyword search | M4 | M4 | Dense vs keyword vs hybrid (alpha) |
| Reranking (cross-encoder) | M4 | M4 | BGE-reranker, Cohere, ColBERT |
| LangChain and LlamaIndex retrievers | M4 | M4 | EnsembleRetriever, VectorIndexRetriever, ParentDocument |
| Parent-child retrieval | M4 | M4 | retrieval.parent-child node |
| Query rewriting and intent | M4 | M4 | query.rewrite, query.intent / model.intent |
| Filters in retrieval | M4, M9 | M4 + M9 | Hard-filters as business guardrail (deeper than IBM) |
| Multi-index routing | M4 | M4 | store.multi-index + retrieval.router; template 07 telecom |
| GraphRAG (Neo4j) | M4 | M4 | Extra depth vs IBM; store.neo4j, retrieval.graph |
| Gradio to evaluate retrieval | M9 | M9 | UI to inspect retrieved chunks |
Course 5 · Multimodal
IBM topics: Whisper, DALL·E, Sora, Hugging Face, watsonx, Granite.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| Multimodal concepts (text/voice/image/video) | M10 | ../10-multimodal/ |
Challenges: alignment, cost, latency |
| STT / Whisper | M10, M9 | M10 + M9 | io.stt node; template 07 call center |
| Vision (describe images, tables→JSON) | M10, M2 | M10 + M2 | model.vision, loader.multimodal; templates 04, 08 |
| Image generation (DALL·E, SDXL) | M10 | M10 | Conceptual + lab |
| Audio generation / TTS | M10 | M10 | Sora mentioned as generative video reference |
| Open models (HF, watsonx, Granite, Llama) | M10, M1 | M10 + M1 | Closed vs open-weights provider comparison |
| Multimodal embeddings and vector DB | M10, M3 | M10 + M3 | Indexing visual descriptions in RAG pipeline |
Course 6 · Fundamentals of AI Agents
IBM topics: tool calling, chaining, LangChain built-in agents, visualization and SQL.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| From RAG to agent (when to agentify) | M6 | ../06-agentes-i/ |
Agentic RAG: the agent decides when to retrieve |
| Tool calling and schemas | M6 | M6 | Tool Calling skill; tool.* nodes |
| Tool chaining | M6 | M6 | Multi-step sequences before responding |
| ReAct loop (reason → act → observe) | M6 | M6 | agent.react; scratch + LangGraph |
| LangChain built-in agents (data, SQL) | M6 | M6 | SQL and data agent in workshop |
| Data visualization agent | M6 | M6 | Explicit lab (IBM mentions it; here with expected result) |
RAG as tool (tool.retriever) |
M6 | M6 | PolicyRAG in template 01 airline |
LangGraph StateGraph (intro) |
M6 | M6 | Conceptual; depth in M7 |
Course 7 · Agentic AI with LangChain / LangGraph
IBM topics: memory, Reflection/Reflexion, ReAct, multi-agent, agentic RAG.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| Memory (short/long term, conversational) | M6, M7 | M6 + ../07-agentes-ii/ |
State in LangGraph; checkpoints in M7 |
| Reflection / Reflexion (self-improvement) | M6 | M6 | ReAct vs Plan-and-Execute vs Reflexion comparison |
| ReAct in depth | M6 | M6 | Workshop: memory + 2 tools, expected sequence |
| Agentic RAG | M6, M4 | M6 + M4 | Query routing by the agent, not fixed pipeline |
| LangGraph: graphs, edges, state | M6, M7 | M6 + M7 | LangChain/LangGraph skill |
| Multi-agent (intro and patterns) | M7 | M7 | Supervisor, hierarchical, collaborative |
| Conditional edges and checkpoints | M7 | M7 | State persistence between steps |
Course 8 · Agentic AI with CrewAI / AutoGen / BeeAI
IBM topics: alternative frameworks and multi-agent patterns.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| CrewAI (agents, tasks, crews, tools) | M7 | ../07-agentes-ii/ |
Hands-on same problem in CrewAI and LangGraph |
| AutoGen / AG2 (conversation between agents) | M7 | M7 | Multi-role dialogue pattern |
| BeeAI | M7 | M7 | Fourth framework in the comparison |
| Multi-agent patterns (supervisor, fan-out) | M7 | M7 | agent.fanout; template 10 logistics |
| When multi-agent vs single agent | M7, M11 | M7 + M11 | Trade-offs table; anti-patterns in M11 |
| Framework selection and combination | M7 | M7 | vs Semantic Kernel (reference) |
| GenAI Agents / Agentic skill | M6, M7 | M6 + M7 | Simple agents → complex orchestration |
Course 9 · Model Context Protocol (MCP)
IBM topics: FastMCP, server/client, STDIO/HTTP, security.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| MCP architecture vs traditional APIs | M8 | ../08-mcp/ |
MCP skill |
| FastMCP: build server (tools, resources, prompts) | M8 | M8 | tool.mcp node |
| MCP client (STDIO and Streamable HTTP) | M8 | M8 | Connect to one and multiple servers |
| MCP security (sampling, roots, permissions) | M8, M9 | M8 + M9 | Permission-based approval; links to AI Security |
| MCP vs proprietary plugins | M8 | M8 | Comparison in guide |
| Integrative workshop | M8 | M8 | Airline PolicyRAG as MCP server |
Course 10 · Capstone
IBM topics: data→deploy, unstructured→JSON, multimodal + multi-agent, MCP, testing.
| IBM topic | Module(s) | Folder | Notes |
|---|---|---|---|
| Full data → production pipeline | M11, M9 | ../11-capstone/ + M9 |
4 deployment targets: chat, Kafka worker, Temporal, batch |
| Unstructured → structured JSON | M2, M5 | M2 + M5 | loader.multimodal, logic.structured |
| Integrated multimodal + multi-agent | M10, M7, M11 | M10 + M7 + M11 | Real templates, not generic dataset |
| MCP in final project | M8, M11 | M8 + M11 | Optional in capstone design |
| AI system testing | M5, M11 | M5 + M11 | RAGAS as test; system testing in M11 |
| Capstone project | M11 | M11 | Beyond IBM: 3 templates from scratch + new design + 50-question exam |
| The 10 industry templates | M1–M11 | All | Map in plantillas-mapeadas.md and PLAN.md §8 |
Visual summary (IBM course → primary module)
| # | IBM course | Primary module(s) |
|---|---|---|
| 1 | Develop GenAI Apps | M1, M5, M9 |
| 2 | Build RAG Apps | M1, M2, M4, M9 |
| 3 | Vector DBs for RAG | M3 |
| 4 | Advanced RAG | M3, M4 |
| 5 | Multimodal | M10 |
| 6 | Fundamentals of AI Agents | M6 |
| 7 | Agentic AI LangChain/LangGraph | M6, M7 |
| 8 | Agentic AI CrewAI/AutoGen/BeeAI | M7 |
| 9 | MCP | M8 |
| 10 | Capstone | M11 (+ cross-cutting M9) |
3. IBM skills → where they are acquired
The IBM program certifies competencies in 11 areas. This table indicates the primary module, reinforcement modules, and the deliverable where the skill is demonstrated.
| IBM skill | Primary module | Also in | How it is demonstrated |
|---|---|---|---|
| RAG | M1 → M5 | M2, M3, M4, M6 | M1 workshop (minimal RAG); full pipeline M2–M5; agentic RAG M6 |
| Vector DBs | M3 | M4 | Chroma + FAISS workshop; metadata filters; pgvector in template 02 |
| Prompt Patterns | M1 | M5, M6 | Few-shot, CoT, system prompts; agent prompts in M6 |
| LangChain / LangGraph | M5, M6, M7 | M1, M2, M4 | LCEL, ReAct, StateGraph, multi-agent, checkpoints |
| OpenAI API | M1 | M3, M5, M6 | Multi-provider interface; embeddings and chat completions |
| Tool Calling | M6 | M7, M8 | ReAct with tools; MCP as evolution of tool calling |
| GenAI Agents | M6 | M7 | ReAct agent + built-in SQL/viz |
| Agentic | M6, M7 | M11 | Reflection, memory, multi-agent, fan-out, capstone |
| Multimodal | M10 | M2, M9 | STT, vision, generation; io.stt in call center |
| MCP | M8 | M9, M11 | FastMCP server + client with permissions |
| AI Security | M9 | M6, M8 | Injection, jailbreak, PII, guardrails, MCP permissions |
Cross-cutting skill not explicitly listed in IBM but covered here:
| Additional skill | Module | Note |
|---|---|---|
| LlamaIndex | M2, M4, M5 | Readers, retrievers, query engines |
| RAG/Agent evaluation | M5, M11 | RAGAS, faithfulness, cost, latency |
| Architecture design | M11 | Flow IR, contracts, anti-patterns |
| Production / SRE | M9 | Idempotency, HITL, Kafka, Temporal, OTel |
4. Extras this course adds (and the IBM program does not emphasize)
IBM covers RAG and agent fundamentals well with frameworks. This course adds production engineering and systematic design capabilities that the Coursera program barely touches.
Real production
| Topic | Module | RAGorbit node(s) | Why it matters |
|---|---|---|---|
| Idempotency | M9 | guardrail.idempotency |
Payments and reservations are not charged twice (Stripe-like) |
| Confirm-gate | M9 | guardrail.confirm |
Irreversible actions require user confirmation |
| Circuit breaker / retry / fallback | M9 | guardrail.resilience |
Graceful degradation when external APIs fail |
| HITL (hardcoded, not LLM) | M9 | hitl.escalate |
Deterministic escalation in healthcare, legal, maintenance |
| Audit to Kafka | M9 | observability.audit |
Regulatory traceability of every tool call |
| Feedback loop | M9 | observability.feedback |
Continuous reranker improvement |
| Temporal / exactly-once | M9 | io.trigger, io.event-source |
Durable workflows; Kafka exactlyOnce |
| 4 deployment targets | M9 | io.* |
chat-service, event-worker, batch, temporal |
| AI Security | M9 | guardrail.* + guides |
Injection, jailbreak, PII leakage, bias |
| Production UIs | M9 | — | Gradio, Streamlit, Flask to operate the system |
Contract awareness and design
| Topic | Module | Reference |
|---|---|---|
| Which node can connect to which (Flow IR) | M0, M11 | docs/01-concepts.md |
| 53 node types in 13 categories | All | catalogo-nodos.md |
| Anti-patterns and design checklist | M11 | ../11-capstone/ |
| Deterministic rules vs delegating to the LLM | M5, M9 | logic.rules, business thresholds |
Ten complete industry cases
Templates in examples/ — each spans several modules:
| Template | Industry | Dominant modules |
|---|---|---|
| 09-RRHH | HR | M1 → M3 |
| 02-Banca | Banking | M2 → M5 |
| 03-Salud | Healthcare | M4 → M9 |
| 04-Seguros | Insurance | M2 → M5, M10 |
| 05-Legal | Legal | M4 |
| 06-Retail | Retail | M6 |
| 07-Telecom | Telecom | M4 → M7, M10 |
| 08-Manufactura | Manufacturing | M2 → M4, M10 |
| 01-Aerolínea | Airline | M6 → M8, M9, M11 |
| 10-Logística | Logistics | M7 → M9, M11 |
Rebuild from scratch (not just use frameworks)
| Layer | What it implies | Where |
|---|---|---|
| ① Design | Why / when / alternatives | All guides |
| ② Scratch | Python stdlib, deterministic mocks | All lab/solucion_scratch.py |
| ③ Framework | LangChain, LangGraph, LlamaIndex, CrewAI… | lab/solucion_framework.py |
| Capstone | Rebuild 09 → 02 → 01 without looking at codegen | M11 |
Advanced RAG with more depth than IBM
| Topic | Module | IBM | This course |
|---|---|---|---|
| GraphRAG / Neo4j | M4 | Light mention | Workshop with store.neo4j + retrieval.graph |
| Multi-index routing | M4 | Basic | store.multi-index + retrieval.router; telecom template |
| Hard-filters as guardrail | M4, M9 | Optional filters | Mandatory WHERE clauses; no filter ⇒ noise demonstrated in lab |
| Parent-child + rerank | M4 | Partial | Legal/medical workshop with precision expected result |
5. Checklist: if you come from the IBM program, start with…
Quick equivalence map so you do not repeat what you already master and can find your place in this curriculum.
You already completed IBM courses 1–2 (GenAI + basic RAG)
- Skip repeated M1 theory and go straight to the workshop in
../01-fundamentos/lab/— validate that you can do minimal RAG in scratch. - Study M2 (
../02-ingesta/): chunking by clause and metadata — IBM does not go as deep on metadata for filters. - Review
catalogo-nodos.md§2–§3 to connect loaders/ingest with RAGorbit.
You already completed IBM courses 3–4 (Vector DBs + Advanced RAG)
- M3 (
../03-embeddings-y-stores/): confirm Chroma ops + FAISS scratch. - Prioritize M4 (
../04-retrieval-y-query/): hybrid, rerank, hard-filters, multi-index, GraphRAG — here we go beyond the IBM program. - Open template 05-Legal or 07-Telecom to see routing in action.
You already completed IBM course 5 (Multimodal)
- M10 (
../10-multimodal/) — quick review if you already used Whisper. - Connect with M2:
loader.multimodaland template 08-Manufactura.
You already completed IBM courses 6–8 (Agents)
- M6 (
../06-agentes-i/): validate ReAct + memory + agentic RAG in scratch. - M7 is mandatory (
../07-agentes-ii/): same problem in CrewAI and LangGraph — IBM separates frameworks; here we compare them in a single workshop. - Template 01-Aerolínea to see agent + tools + guardrails.
You already completed IBM course 9 (MCP)
- M8 (
../08-mcp/): server + client + security/permissions — goes deeper than the typical Coursera lab. - Link with M9: MCP permissions + AI Security.
You already completed IBM capstone 10
- Do not skip M9 (
../09-produccion-y-seguridad/): idempotency, HITL, Kafka, Temporal — probably your biggest gap vs this course. - M11 (
../11-capstone/): rebuild templates 09, 02, 01 without looking at solutions; design new architecture; integrative exam.
Quick table IBM → first module to open
| If you master IBM course… | Start in this module | If it is too easy, skip to… |
|---|---|---|
| 1 Develop GenAI Apps | M2 or M1 workshop | M5 (JSON/eval) |
| 2 Build RAG Apps | M3 | M4 |
| 3 Vector DBs | M4 | M5 |
| 4 Advanced RAG | M5 or M6 | M9 (hard-filters in prod) |
| 5 Multimodal | M6 | M9 |
| 6 Fundamentals Agents | M7 | M8 |
| 7 LangGraph Agentic | M7 (CrewAI/AutoGen) | M8 |
| 8 CrewAI/AutoGen/BeeAI | M8 | M9 |
| 9 MCP | M9 | M11 |
| 10 Capstone | M9 + M11 | — |
Common setup (all profiles)
- M0 (
../00-setup/): environment, Python review, firstflow.jsonin RAGorbit. - Read
HANDOFF.md §3: tri-modal method and scratch/stdlib constraint.
6. Conclusion
| Criterion | IBM Coursera program | This course (M0–M11) |
|---|---|---|
| RAG + Agents + MCP + Multimodal syllabus | ✅ 10 courses | ✅ 100% covered (table §2) |
| IBM certifiable skills (11 areas) | ✅ | ✅ All mapped (table §3) |
| Frameworks (LangChain, LangGraph, LlamaIndex, CrewAI, AutoGen, BeeAI, MCP) | ✅ | ✅ + comparisons and trade-offs |
| Production (idempotency, HITL, audit, Temporal) | ⚠️ Minimal | ✅ Dedicated M9 |
| Design and contracts (Flow IR, 53 nodes) | ❌ | ✅ M0 + M11 + catalog |
| Real industry cases | ⚠️ Generic | ✅ 10 templates |
| Implementation from scratch | ⚠️ Partial | ✅ Scratch layer in every workshop |
| GraphRAG, multi-index, hard-filters | ⚠️ Superficial | ✅ M4 in depth |
| Final evaluation | Coursera project | ✅ Triple capstone (rebuild + design + exam) |
Verdict: plan v2 (PLAN.md §11) fully covers the IBM "RAG and Agentic AI" syllabus and goes beyond it in production, design awareness, industry cases, and ability to rebuild systems from scratch. The IBM syllabus served as a safety net for coverage; the material, pedagogical order, and deliverables are this repository's own.
Related documents: PLAN.md · HANDOFF.md · catalogo-nodos.md · plantillas-mapeadas.md · tecnologias-comparadas.md