---

name: indexing description: "Initialize GraphRAG workspaces and run, update, inspect, or troubleshoot indexing pipelines, including standard vs fast methods, workflow selection, output tables, BYOG indexes, embeddings, cache/reporting/storage, and incremental update failures." disable-model-invocation: true metadata: disco-role: operating license: MIT

GraphRAG Indexing

Use this sub-skill when a task mentions graphrag init, graphrag index, graphrag update, build_index, indexing methods, workflows, output parquet tables, BYOG tables, embeddings generated during indexing, or update failures.

Route First

• For config schema details, input providers, model/env setup, and storage provider configuration, use ../configuration-data/.
• For querying a completed index with global, local, DRIFT, or basic search, use ../querying/.
• For generating or editing extraction/community report prompts, use ../prompt-tuning/.
• For implementing custom factories/providers beyond workflow selection, use ../package-extensions/.

Common Tasks

Initialize a Workspace

graphrag init --root <project-root>

This creates a default config and prompt layout. After initialization, add input documents under the configured input location, set model credentials outside committed config when needed, and dry-run the indexer before a costly run.

Dry-Run and Build an Index

graphrag index --root <project-root> --dry-run
graphrag index --root <project-root> --method standard

standard is the default. Use --method fast for FastGraphRAG when lower cost matters more than rich LLM-derived entity and relationship descriptions. Add --verbose while debugging and --no-cache only when cached LLM responses must be ignored.

Incrementally Update an Index

graphrag update --root <project-root> --method standard

update calls the same indexing API with is_update_run=True, copies previous output tables into update storage, builds a delta index, then runs update merge workflows. Confirm previous outputs exist before updating, and verify that changed, deleted, duplicate, and unchanged document titles behave as expected for the project.

Call the Python API

import asyncio
from graphrag.api import build_index
from graphrag.config.load_config import load_config

config = load_config(root_dir="<project-root>")
results = asyncio.run(build_index(config, method="standard", verbose=True))
if any(result.error for result in results):
    raise RuntimeError([result.workflow for result in results if result.error])

The verified signature is build_index(config, method='standard', is_update_run=False, callbacks=None, additional_context=None, verbose=False, input_documents=None). input_documents may be a pandas DataFrame with GraphRAG document columns to bypass file loading.

Method Choice

• standard: LLM graph extraction, relationship extraction, description summarization, optional claims, community reports, and embeddings. Prefer for high-fidelity graph outputs.
• fast: NLP noun phrase extraction and co-occurrence relationships, LLM community reporting, and embeddings. Prefer for lower cost or large exploratory runs; check SpaCy/NLP model requirements when using non-default extractors.
• standard-update / fast-update: internal pipeline names selected when build_index(..., is_update_run=True) or graphrag update is used with the base method.
• Custom workflows: a config workflows list overrides built-in method pipelines; use it for BYOG or partial reruns after confirming source tables exist.

Outputs to Expect

Default file table output is parquet. Core tables are documents, text_units, entities, relationships, communities, and community_reports; covariates appears when claim extraction is enabled. Embeddings are written to the configured vector store for enabled embedding names such as text_unit_text, entity_description, and community_full_content.

Use bundled helpers:

python sub-skills/indexing/scripts/inspect_index_config.py --root <project-root>
python sub-skills/indexing/scripts/validate_index_outputs.py --output <project-root>/output --use global
python sub-skills/indexing/scripts/run_mock_index_workflow.py

BYOG Pattern

For bring-your-own graph, place entities.parquet and relationships.parquet in the configured output table location before running only the downstream workflows. Add text_units.parquet and generate_text_embeddings when local, DRIFT, or basic search will need chunks or vectors.

Minimal global-search BYOG workflow:

workflows: [create_communities, create_community_reports]

Expanded BYOG workflow for non-global search:

workflows: [create_communities, create_community_reports, generate_text_embeddings]

FastGraphRAG-style BYOG reporting can use create_community_reports_text when entities and relationships lack descriptions but have valid text_unit_ids links.

References

• See references/indexing-api-cli.md for CLI/API entry points and lifecycle details.
• See references/indexing-config-and-data.md for config, storage, input, BYOG, cache, and vector considerations.
• See references/indexing-workflows-and-outputs.md for built-in workflow order and output table expectations.
• See references/troubleshooting.md for common failures and targeted fixes.