Agent Skills: LlamaIndex Knowledge Patch

LlamaIndex Knowledge Patch

Use this skill when writing, reviewing, migrating, or debugging Python code that uses LlamaIndex core, ingestion, storage, retrieval, agents, workflows, or its separately versioned integrations.

Inspect the application's installed distributions and lockfile before applying package-specific advice. The umbrella package, core, workflows, and integrations do not necessarily share a distribution version.

Reference Index

| Reference | Topics | | --- | --- | | Agents and workflows | Agent selection, handlers, memory, events, concurrency, state, resources, validation, AG-UI | | Ingestion and storage | Pipelines, caches, document updates, persistence, retrieval, parsers, graph extraction, vector stores | | Migration and packaging | Core API migration, dependency coordination, Python and Pydantic changes, model and tool integrations |

First Checks

Before changing code:

1. Inspect the versions of llama-index, llama-index-core, and every llama-index-* integration actually imported by the application.
2. Preserve and review the resolved lockfile; never infer integration versions from the core version.
3. Identify whether persistence includes local stores, a remote vector store, or both.
4. Determine whether each ingestion run is a complete source inventory or only a partial crawl before enabling deletion behavior.
5. Treat agent and workflow runs as asynchronous, streamable executions.
6. Re-run ingestion, retrieval, agent, and workflow integration tests after an upgrade.

Breaking Migration Priorities

Replace removed configuration abstractions

Do not construct ServiceContext or LLMPredictor. Set process-wide defaults through Settings, or pass LLMs, embedding models, and transformations directly when multiple configurations must coexist.

from llama_index.core import Settings

Settings.llm = llm
Settings.embed_model = embed_model

Audit custom readers, nodes, parsers, tools, models, validators, serialization, and integrations for Pydantic v2 behavior.

Coordinate package upgrades

Upgrade a coherent dependency set rather than changing core alone or forcing every distribution to the same version. Respect each distribution's metadata, including its Python floor, and preserve the resolved environment.

from importlib.metadata import version

core_version = version("llama-index-core")
starter_version = version("llama-index")

Redesign obsolete agent code

Use the workflow-based agent imports. Older agent, runner, and worker examples do not describe the current execution model.

from llama_index.core.agent.workflow import (
    AgentWorkflow,
    FunctionAgent,
    ReActAgent,
)

Choose FunctionAgent only when the configured LLM provides compatible native tool calling. Use ReActAgent for text-based reasoning and action parsing, or CodeActAgent for code-action scenarios.

Redesign pipelines as workflows

A workflow is not a renamed QueryPipeline. Model the control flow with typed events, asynchronous @step methods, branches, loops, Context, streaming, and checkpoint-aware execution.

Use llama_index.core.workflow inside core applications. Standalone workflow applications may use the llama-index-workflows distribution and import from workflows.

Agent and Workflow Quick Reference

Keep and await the handler

agent.run(...) and workflow.run(...) return awaitable handlers. Keep the handler, consume its event stream if needed, and await that same object for the final result.

handler = agent.run("What is 12 times 34?")
async for event in handler.stream_events():
    ...
result = await handler

Create memory explicitly

Create Memory with Memory.from_defaults and pass it to run. Conversation history and memory blocks are distinct from workflow Context, which carries per-run execution state and events.

from llama_index.core.memory import Memory

memory = Memory.from_defaults(session_id="session-123", token_limit=40000)
response = await agent.run("...", memory=memory)

Separate state from resources

Store serializable per-run state through the asynchronous ctx.store get/set/edit interface. Supply clients, indexes, models, and configuration as workflow resources so dependency injection and validation do not mistake live objects for checkpointable state.

Validate event graphs

Call workflow.validate() in tests or during startup to detect missing start/stop paths, events without consumers or producers, and dead ends.

workflow = RagFlow(timeout=60)
workflow.validate()

For dynamic fan-out, emit work with Context.send_event and join it with Context.collect_events. Do not assume results arrive in input order.

Ingestion and Persistence Quick Reference

Own updates through an ingestion pipeline

VectorStoreIndex.from_documents is convenient, but use IngestionPipeline when the application needs transformation caching, document update strategy, or direct vector-store insertion.

Stable document.doc_id and node.ref_doc_id values drive hash-based duplicate, update, and deletion detection. A deletion strategy that removes absent documents is safe only when the run is an authoritative inventory.

Treat caches and document state separately

An ingestion cache keys reuse by the node-plus-transformation combination. Persisted cache state is separate from docstore state; saving one does not save the other.

pipeline.persist("./pipeline-state")
pipeline.load("./pipeline-state")

Ensure custom transformations expose stable serialized settings or hashes so a behavior change cannot silently reuse stale output.

Choose the correct execution path

Use pipeline.arun() in asynchronous code. num_workers enables process-based parallelism for suitable synchronous transformations; provider clients and non-picklable transformations may make it invalid or slower.

nodes = await pipeline.arun(documents=documents, num_workers=4)

Verify the whole persistence boundary

index.storage_context.persist() saves local stores, not necessarily the external vector collection. A remote collection alone can also omit node text, mappings, or document state. Confirm that both sides exist, use a compatible embedding configuration, and select the expected index ID when a directory contains multiple indexes.

from llama_index.core import StorageContext, load_index_from_storage

index.storage_context.persist(persist_dir="./storage")
storage_context = StorageContext.from_defaults(persist_dir="./storage")
index = load_index_from_storage(storage_context)

Rebuild or explicitly verify persisted vector indexes whenever the embedding model or its configuration changes.

Upgrade Regression Checklist

After changing packages or configuration, verify:

• all derived nodes for a re-ingested document are present;
• zero-valued retrieval thresholds remain effective;
• omitted metadata-filter conjunctions behave as AND;
• tool schemas exclude variadic parameters from required fields;
• structured agent outputs honor the configured callback and output class;
• streaming memory writes retain every block and populate response text;
• citation fragments have distinct identities and usable offsets;
• parser node counts and boundaries match the application's expectations;
• remote metadata retains falsy values and pre-existing properties;
• asynchronous reranking, graph extraction errors, and multimodal input follow the application's chosen behavior.

Consult the topic references for the exact package releases and edge cases behind these checks.