AutoGraph RAG Strategizer Reference

Trigger RAG Strategizer

Analyze existing clusters and assign a RAG strategy to each of them. Must be called after a corpus build is completed.

Recommended path: Call after the corpus build reaches completed, and before POST /v1/orchestrate. If you rebuild the corpus, run the strategizer again so that rags reflects the current clusters.

Request

{
  "project": "my_project",
  "complexity": "high",
  "extract_images_default": false,
  "categories": ["legal", "marketing"],
  "max_parallel_clusters": 5
}

Parameters

How complexity is applied

The percentage applies to the ranked cluster list: the top round(percentage / 100 * cluster_count) clusters get FullGraphRAG and the rest get VectorRAG. With few clusters, the rounding decides the outcome. At moderate, a project with a single cluster gets round(0.5) == 0 FullGraphRAG clusters, so the whole project ends up on VectorRAG:

A VectorRAG partition has no Entities and no Communities, so it cannot serve LOCAL, GLOBAL, or UNIFIED queries. Send very_high when every partition has to support entity-based retrieval. See Retrieval capability per strategy.

In the web interface, the split between the two strategies is exposed as the Complexity slider of the strategy generation wizard, which runs from Vector only (no FullGraphRAG) to Graph + images (FullGraphRAG everywhere, with image extraction). See Web Interface.

Response

{
  "strategize_job_id": "strat_1738000000_a1b2c3d4"
}

The strategizer runs asynchronously. Poll GET /v1/rag-strategizer/jobs/{strategize_job_id} for the live progress.

For each cluster the strategizer samples documents, computes a complexity score, and assigns FullGraphRAG or VectorRAG. For FullGraphRAG clusters it also uses an LLM to generate a domain-specific ontology of 8 to 12 entity types. The ontology and the strategy profiles are persisted to the rags collection and later passed to the Importer to constrain entity extraction, see Per-cluster ontology.

HTTP Example

curl -X POST \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <token>" \
  -d '{
    "project": "my_project",
    "complexity": "high",
    "extract_images_default": false,
    "max_parallel_clusters": 5
  }' \
  https://<EXTERNAL_ENDPOINT>:8529/autograph/v1/rag-strategizer/analyze

Monitor a strategizer job

Poll the live progress of an analyze run with the strategize_job_id that POST /v1/rag-strategizer/analyze returned. Use it to drive a progress bar while the job runs in the background.

Parameters

Response

{
  "status": "running",
  "progress": 42.0,
  "clusters_total": 10,
  "clusters_done": 4
}

Poll until status is completed or failed. A terminal snapshot is kept for a bounded window, one hour by default, and is then evicted. The status lives in memory and is lost on a pod restart.

HTTP Example

curl -H "Authorization: Bearer <token>" \
  https://<EXTERNAL_ENDPOINT>:8529/autograph/v1/rag-strategizer/jobs/strat_1738000000_a1b2c3d4

Retrieve RAG Strategies

Retrieve all RAG strategies that have been created.

Recommended path: Optional audit step after the strategizer job reaches completed. Use it to check which strategy each cluster was assigned before you orchestrate, and to get the rag_partition_id values that POST /v1/graph/recluster takes. Safe to call anytime for read-only inspection.

Request

No body or query parameters.

Response

{
  "strategies": [
    {
      "cluster_id": "cluster_legal_0",
      "strategy_type": "FullGraphRAG",
      "rag_partition_id": "legal_0_a",
      "entity_types": ["CONTRACT", "JURISDICTION", "LEGISLATION"],
      "document_count": 25,
      "parameters": {
        "rag_mode": "FullGraphRAG",
        "module": "legal",
        "batch_size": "100",
        "enable_chunk_embeddings": "true",
        "enable_edge_embeddings": "false",
        "chunk_token_size": "1200",
        "chunk_overlap_token_size": "100",
        "community_report_num_findings": "5-10",
        "enable_semantic_units": "true"
      }
    },
    {
      "cluster_id": "cluster_legal_1",
      "strategy_type": "VectorRAG",
      "rag_partition_id": "legal_1_b",
      "entity_types": [],
      "parameters": {
        "rag_mode": "VectorRAG",
        "module": "legal"
      }
    }
  ],
  "total_strategies": 2,
  "strategy_type_counts": {
    "FullGraphRAG": 1,
    "VectorRAG": 1
  }
}

HTTP Example

curl -H "Authorization: Bearer <token>" \
  https://<EXTERNAL_ENDPOINT>:8529/autograph/v1/rag-strategizer/strategy

Update a cluster strategy

Override the strategy type, the entity types, and the image extraction flag of a single cluster. Use it when the automatically assigned strategy is not what you want.

Recommended path: Call after GET /v1/rag-strategizer/strategy and before POST /v1/orchestrate.

Request

{
  "strategy_type": "FullGraphRAG",
  "entity_types": ["PERSON", "ORGANIZATION", "LOCATION"],
  "extract_images": false
}

Response

{
  "strategy": {
    "cluster_id": "cluster_legal_0",
    "strategy_type": "FullGraphRAG",
    "rag_partition_id": "legal_0_a",
    "entity_types": ["PERSON", "ORGANIZATION", "LOCATION"],
    "document_count": 25,
    "parameters": {
      "rag_mode": "FullGraphRAG",
      "updated_at": "1752844800.0",
      "extract_images_default": "false",
      "enable_semantic_units": "false",
      "process_images": "false",
      "enable_semantic_unit_embeddings": "false",
      "batch_size": "100",
      "enable_chunk_embeddings": "true",
      "chunk_token_size": "1200",
      "chunk_overlap_token_size": "100"
    }
  }
}

HTTP Example

curl -X PATCH \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <token>" \
  -d '{
    "strategy_type": "VectorRAG",
    "entity_types": [],
    "extract_images": false
  }' \
  https://<EXTERNAL_ENDPOINT>:8529/autograph/v1/rag-strategizer/strategy/cluster_legal_0

Retrieval capability per strategy

The strategy of a cluster decides what a query against that partition can answer. This is by design: a VectorRAG partition has no Entities and no Communities, and the entity-based query modes are built on that entity layer.

The partition is the unit, not the project. A project that holds both strategies answers entity-based queries only from its FullGraphRAG partitions. A question whose answer sits in the documents of a VectorRAG partition still comes back unanswered, because the partition cannot serve the mode.

To make every partition entity-capable, send complexity: "very_high" to POST /v1/rag-strategizer/analyze, or override individual clusters with PATCH /v1/rag-strategizer/strategy/{cluster_id}. Do both before you orchestrate.

Next Steps

• Orchestrate Pipeline: Automatically build knowledge graphs for all strategies
• Per-cluster ontology: How entity types shape your knowledge graph