AI Builder

6 skills inclus

• 01

  RAG Pipeline

  Fait construire un RAG qui répond à partir de tes documents avec citations : découpage, embeddings, base vectorielle, reranking et évaluation.+

  sf-rag-pipeline/SKILL.md, de wshobson/agents, licence MIT.

  1---

  2name: sf-rag-pipeline

  3description: "Builds retrieval-augmented generation (RAG) so an LLM answers from your documents with citation…

  4---

  5

  6# RAG Pipeline

  7

  8Ground an LLM's answers in your own documents. A RAG system is only as good as its retrieval: measure retriev…

  9

  10## When to use

  11

  12- Q&A or chat over internal docs, product documentation, tickets, contracts, code.

  13- An existing assistant hallucinates, misses obvious documents or cites the wrong source.

  14

  15Not for: data that fits in the prompt (just include it), structured questions better answered by SQL, or teac…

  16
• 02

  Choosing Embeddings

  Fait choisir et tester le modèle d'embeddings sur tes propres requêtes, puis régler dimensions, découpage, cache et plan de ré-indexation.+

  sf-embedding-choice/SKILL.md, de wshobson/agents, licence MIT.

  1---

  2name: sf-embedding-choice

  3description: "Chooses, benchmarks and tunes the embedding model behind semantic search or RAG: compares hoste…

  4---

  5

  6# Choosing Embeddings

  7

  8The embedding model decides what "similar" means for your search. Pick it with a small benchmark on your own…

  9

  10## When to use

  11

  12- Before the first indexing of a RAG or semantic search corpus.

  13- Retrieval returns topically close but wrong results.

  14- Storage or latency costs of vectors are too high.

  15- New content type: source code, several languages, legal or financial text.

  16
• 03

  Hybrid Search

  Fait combiner recherche vectorielle et recherche par mots-clés avec fusion RRF et reranking, pour retrouver noms, codes et termes rares.+

  sf-hybrid-search/SKILL.md, de wshobson/agents, licence MIT.

  1---

  2name: sf-hybrid-search

  3description: "Combines vector (semantic) search with keyword search (BM25, Postgres full-text, Elasticsearch/…

  4---

  5

  6# Hybrid Search

  7

  8Vectors find meaning; keywords find exact strings. Run both, fuse the ranked lists, then rerank. Hybrid searc…

  9

  10## When to use

  11

  12- Queries include identifiers, SKUs, error codes, function names, people or product names, acronyms.

  13- Domain vocabulary the embedding model handles poorly.

  14- Measured recall with vector search alone is below target.

  15

  16Not needed when queries are purely conversational and dense retrieval already meets the evaluation target.
• 04

  LLM Evaluation

  Fait mettre en place des évaluations pour une fonction LLM : jeu de test versionné, contrôles exacts, juge LLM calibré et blocage des régressions en CI.+

  sf-llm-evals/SKILL.md, de wshobson/agents, licence MIT.

  1---

  2name: sf-llm-evals

  3description: "Sets up evaluation for an LLM feature so prompt, model and retrieval changes are measured inste…

  4---

  5

  6# LLM Evaluation

  7

  8An LLM feature without evals is untested code. Build the smallest evaluation that can tell a better version f…

  9

  10## When to use

  11

  12- Before shipping a prompt, model, retrieval or agent change.

  13- Comparing models or prompts; choosing a cheaper model.

  14- After a production complaint, to capture the failure as a test.

  15

  16## Process
• 05

  LangGraph Agents

  Fait concevoir et déboguer des agents LangGraph : état typé avec reducers, nœuds, arêtes, Command, Send, streaming et gestion d'erreurs.+

  sf-langgraph-agents/SKILL.md, de langchain-ai/langchain-skills, licence MIT.

  1---

  2name: sf-langgraph-agents

  3description: "Builds and debugs LangGraph agent workflows in Python or TypeScript: designing the graph from a…

  4---

  5

  6# LangGraph Agents

  7

  8LangGraph models an agent or workflow as a directed graph of nodes over shared state. Design the flowchart fi…

  9

  10## When to use

  11

  12- Writing, changing or debugging LangGraph code in Python or TypeScript.

  13- Building a multi-step agent with branches, loops, tools, parallel work or persistence.

  14

  15Not for: a single LLM call or a fixed chain (call the model directly), or approval and pause/resume flows on…

  16
• 06

  LangGraph Human-in-the-Loop

  Fait ajouter des validations humaines à un agent LangGraph : pause avant une action risquée, reprise, boucle de validation, effets de bord idempotents.+

  sf-langgraph-approvals/SKILL.md, de langchain-ai/langchain-skills, licence MIT.

  1---

  2name: sf-langgraph-approvals

  3description: "Adds human-in-the-loop steps to LangGraph agents: pausing with interrupt() before risky actions…

  4---

  5

  6# LangGraph Human-in-the-Loop

  7

  8Pause a graph, show a person what is about to happen, and continue with their decision. The pause survives re…

  9

  10## When to use

  11

  12- An agent is about to send, pay, delete, deploy or publish: require approval.

  13- A tool call or generated content needs review or editing before use.

  14- The graph needs information only the user can give.

  15

  16Graph design, reducers, edges and retries in general: `sf-langgraph-agents`.

Extraits. Les dossiers complets sont livrés après le paiement.

Installation