llamaindex
Skill172 starsChanged 4 months agoMITView on GitHub
Build LLM applications with the LlamaIndex framework. Use when working with LlamaIndex or comparing RAG and agent orchestration frameworks. Do not use this skill for unrelated requests; route to the nearest named specialist.
3 other projects use a version of this.
Adopted by
Most recently taken up by
- G1Joshi/Agent-Skillstheir own version, first seen Oct 13, 2026
- mhiskall282/Securing-the-Digital-Mine-UNESCO-Projecttheir own version, first seen Sep 20, 2026
- Clay-HHK/claude-skillstheir own version, first seen Mar 5, 2026
Other versions3 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 3
- G1Joshi/Agent-Skills★ 15same name
- Clay-HHK/claude-skills★ 1same name
- mhiskall282/Securing-the-Digital-Mine-UNESCO-Project★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository192 other files it gives its agents.
- agent-production-operations/AGENTS.mdAGENTS.md
- AGENTS.mdAGENTS.md
- forward-deployed-engineering/AGENTS.mdAGENTS.md
- production-excellence/AGENTS.mdAGENTS.md
- product-lifecycle/AGENTS.mdAGENTS.md
- tailscale/AGENTS.mdAGENTS.md
- workflow-architect/AGENTS.mdAGENTS.md
- llms.txtllms.txt
- actuarial-risk-modelingSkill
- adr-authoringSkill
- agent-councilSkill
- agent-evals-and-observabilitySkill
- agent-production-operationsSkill
- agent-skillsSkill
- ai-governanceSkill
- ai-operating-economicsSkill
- amazon-memoSkill
- analog-occultismSkill
- anydocSkill
- api-design-and-evolutionSkill
- artifact-pyramidsSkill
- ascii-city-engineSkill
- autogenSkill
- backend-engineeringSkill
- binary-analysisSkill
- bmadSkill
- brand-designerSkill
- c4-diagrammingSkill
- camera-hackingSkill
- capacity-and-cost-engineeringSkill
- chief-of-staff-methodologySkill
- cli-builderSkill
- cncf-landscapeSkill
- code-reviewSkill
- color-managementSkill
- community-guidesSkill
- conditional-customer-successSkill
- confluence-cliSkill
- constrained-optimizationSkill
- crewaiSkill
- crmSkill
- crowdsecSkill
- cryptpadSkill
- cyberpunkSkill
- daily-life-discoverySkill
- data-architectSkill
- data-cleaningSkill
- data-engineeringSkill
- data-scientistSkill
- de-spinSkill
- digital-twinSkill
- docker-composeSkill
- documentsSkill
- dsm5Skill
- dspySkill
- electronicsSkill
- emailSkill
- enterprise-architectureSkill
- epubSkill
- esp32-developmentSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
1 rule. 0 are common, 1 is distinctive, and 0 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
1
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 129
When something goes wrong, find your symptom and follow the recovery path:
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
LlamaIndex Expert Skill
LlamaIndex is an MIT-licensed Python framework for building LLM applications over your data. In 2026, it has evolved from a RAG indexing library into an event-driven workflow framework with integrated production runtime (llama-deploy), agent orchestration (AgentWorkflow), knowledge graph construction (PropertyGraphIndex), and OpenTelemetry-native observability.
The framework is organized around seven core primitives: Reader (data loaders), Document/Node (chunked content model), Index (data structures over Nodes), Retriever (relevant Node selection), Query Engine (retriever + synthesis), Agent (LLM with tools), and Workflow (event-driven orchestration).
Key Principles
These principles govern every decision when building with LlamaIndex. Read them before proceeding to the reference guides.
- Decouple retrieval chunks from synthesis chunks. The embedding representation that retrieves well differs from the context representation that generates well. Use
SentenceWindowNodeParser+MetadataReplacementNodePostProcessorfor this pattern. - Rerank before you generate. Hybrid retrieval + reranker is the minimum viable production RAG configuration.
- Agents are Workflows.
FunctionAgentandAgentWorkfloware pre-configured Workflows. Drop to rawWorkflowwhen you need custom control flow. - Graphs are not just vector stores.
PropertyGraphIndexadds structural path traversal that vector similarity cannot provide — combine both for maximum retrieval quality. - Evaluate in the same process. Span-attached evaluation preserves the connection between the output and the retrieval context that produced it.
Where to Start
The pipeline has 9 phases from Ingest to Deploy. If you're joining mid-stream with existing work, find your entry point:
| You already have... | Start at phase | What to do |
| Nothing — blank project | Ingest | Set up data loading, then proceed through the full pipeline |
| Documents in a directory | Chunk | Choose a chunking strategy, build your index |
| A working vector index | Retrieve | Add hybrid search, reranking, metadata filters |
| An existing RAG pipeline to harden | Deploy | Add observability, llama-deploy, production debugging |
| A need to measure and improve quality | Evaluate | Set up evaluators, ParamTuner, span-attached scoring |
| Nothing — comparing frameworks | See Framework Routing Guide | Don't start the pipeline — pick the right tool first |
Pipeline Mode
Different tasks need different levels of rigor. Match your scope to a mode:
| Mode | When | Phases to run | Skip |
| Quick | Single query, one source, exploration | Ingest → Chunk → Index → Retrieve | Reranking, metadata filters, observability, evaluation |
| Full | Production RAG, multiple sources, compliance | Ingest → Chunk → Index → Retrieve → Agent/Workflow → Deploy → Evaluate | Nothing — run all phases |
| Evaluate | Benchmarking, regression testing | Ingest → Chunk → Index → Evaluate | Retrieve, Agent, Workflow, Deploy (run offline) |
| Graph | Knowledge graph construction | Ingest → Chunk → Graph → Retrieve | Agent, Workflow, Deploy (query via graph index directly) |
Rule of thumb: if you're shipping to users, run Full mode. If you're exploring, run Quick. If you're measuring, run Evaluate.
Quick Reference
| Phase | Task | Approach | Reference |
| Ingest | Load data | SimpleDirectoryReader("./data").load_data() | references/architecture.md |
| Chunk | Parse documents into nodes | SentenceSplitter(chunk_size=1024) | references/rag-strategies.md |
| Index | Build vector index | VectorStoreIndex.from_documents(docs) | references/rag-strategies.md |
| Retrieve | Hybrid search + rerank | BM25Retriever + CohereRerank | references/rag-strategies.md |
| Agent | Multi-agent orchestration | AgentWorkflow(agents=[...]) | references/agent-patterns.md |
| Workflow | Event-driven pipeline | class MyFlow(Workflow): @step | references/workflows.md |
| Graph | Knowledge graph | PropertyGraphIndex.from_documents(docs) | references/property-graph-index.md |
| Evaluate | RAG evaluation | FaithfulnessEvaluator().evaluate_response(...) | references/evaluation-observability.md |
| Deploy | Production deployment | deploy_workflow(workflow=MyFlow()) | references/production-deployment.md |
When to Use This Skill
Load this skill any time you are:
- Building a RAG pipeline over enterprise or personal data
- Comparing LlamaIndex with LangChain, Haystack, or DSPy
- Designing multi-agent systems with handoff between specialist agents
- Deploying an LLM application to production with observability
- Constructing knowledge graphs from unstructured documents
- Debugging common LlamaIndex failures (retrieval miss, handoff bug, async issues)
When NOT to Use LlamaIndex — Framework Routing Guide
This skill is part of a portfolio of framework skills. When deciding which framework fits, use this routing table:
| Scenario | Reach for | Why |
| I have documents I need to query | LlamaIndex | Data ingestion, hybrid retrieval, reranking, and knowledge graphs are first-class primitives |
| I have agents I need to orchestrate | LangGraph | State-machine semantics, time-travel debugging, and human-in-the-loop pauses are the core design |
| I have a tool I need to wrap as an agent | PydanticAI | Type-safe agent definitions with dependency injection, minimal abstraction over LLM calls |
| Data-heavy RAG over PDFs, SQL, Slack, 200+ sources | LlamaIndex | LlamaHub connectors, LlamaParse for documents, hybrid retrieval out of the box |
| Complex multi-agent state machines with checkpoints | LangGraph | Graph topology control — supervisor, subgraphs, hierarchical teams, built-in checkpointer |
| Agent-centric app where type safety matters more than data pipelines | PydanticAI | Agents as Pydantic models, DI, structured outputs — the data layer is your code |
| Document parsing quality matters (tables, charts, handwriting) | LlamaIndex | LlamaParse is purpose-built for this |
| Production NLP search pipelines | Haystack | Pipeline composition model is more mature for search-specific workloads |
| Optimization-driven prompt programming | DSPy | Compiled prompt programs, not retrieval pipelines |
Reference Files
| Reference | Load when | File |
| Core Architecture | Understanding the 7 primitives, Settings, data flow | references/architecture.md |
| RAG Strategies | Building RAG pipelines from basic to advanced | references/rag-strategies.md |
| Agent Patterns | Multi-agent orchestration with AgentWorkflow | references/agent-patterns.md |
| Workflows | Event-driven step composition and durable execution | references/workflows.md |
| Production & Deployment | llama-deploy, debugging, failure modes | references/production-deployment.md |
| Property Graph Index | Knowledge graph construction and hybrid retrieval | references/property-graph-index.md |
| Evaluation & Observability | Metrics, tracing, span-attached scoring | references/evaluation-observability.md |
| Integration Ecosystem | Vector stores, LlamaHub, LlamaParse, ecosystem | references/integration-ecosystem.md |
| FAQ & Troubleshooting | Common errors and their fixes | references/faq-and-troubleshooting.md |
| Worked RAG Example | Complete end-to-end pipeline from ingest to deploy | references/example-rag-pipeline.md |
| Evaluation Workflow | ParamTuner, evaluators, batch scoring, best practices | references/evaluation-workflow.md |
Template Files
| Template | When to use | File |
| Basic RAG | Single-source query, getting started | templates/basic-rag.py |
| Agentic RAG | Multi-source data with agent routing | templates/agentic-rag.py |
| Custom Workflow | Custom control flow, branching logic | templates/custom-workflow.py |
| Production Deploy | Wrapping a workflow as a microservice | templates/production-deploy.py |
Scripts
| Script | Purpose | File |
| check-setup | Verify LlamaIndex installation and configuration | scripts/check-setup.py |
Troubleshooting — Structured Recovery Guide
When something goes wrong, find your symptom and follow the recovery path:
Retrieval & Answer Quality
| Symptom | Likely cause | Immediate fix | Permanent fix | Reference |
| Answers are poor or hallucinated | No reranker on hybrid retrieval | Add CohereRerank(top_n=5) as node_postprocessor | Reranking is mandatory for any production RAG | references/rag-strategies.md |
| Retrieval misses obvious content | Default chunking breaks semantics | Switch to SemanticSplitterNodeParser(breakpoint_percentile_threshold=95) | Tune chunk size with ParamTuner | references/rag-strategies.md |
| Wrong tenant's data returned | Missing metadata filters | Add MetadataFilters(filters=[ExactMatchFilter(key="tenant_id", ...)]) | Always wire metadata filters at retriever level | references/rag-strategies.md |
| Only one type of query works well | Single retrieval strategy | Combine BM25 + vector via hybrid retriever | Add RouterQueryEngine for query-type routing | references/rag-strategies.md |
Agent & Workflow Failures
| Symptom | Likely cause | Immediate fix | Permanent fix | Reference |
| Agent waits silently after handoff | AgentWorkflow handoff bug | Extend FunctionAgent.take_step to re-locate last user message | Apply the handoff fix on all production agents | references/agent-patterns.md |
| Workflow doesn't run | Forgot await | Add await before w.run(...) and all step calls | All step methods are async coroutines | references/workflows.md |
| Step executes but result is lost | State not persisted | Use ctx.store.edit_state() for shared state | Only ctx.store survives across steps | references/workflows.md |
| Crash loses all progress | No checkpoint snapshots | Add Context.to_dict() save on step completion | Durable workflows need explicit checkpointing | references/workflows.md |
Deployment & Observability
| Symptom | Likely cause | Immediate fix | Permanent fix | Reference |
| llama-deploy deployed but requests time out | Redis not running | Start redis-server | Redis is mandatory — control plane won't route without it | references/production-deployment.md |
| Spans missing in observability UI | Instrumentation called too late | Move instrument() call before workflow instantiation | Always instrument before creating any Workflow object | references/production-deployment.md |
| Wrong data returned (cross-tenant) | Missing metadata filters | Add tenant filter to all retrievers | Filter at retriever level, not in post-processing | references/production-deployment.md |
Recovery Workflow
For any failure, follow this cycle:
- Identify the symptom from the tables above
- Apply the immediate fix — this gets you running
- Implement the permanent fix — this prevents recurrence
- Verify with evaluation — run
FaithfulnessEvaluatoron a held-out query set - Document the fix — add the root cause to
references/faq-and-troubleshooting.md
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.