Skip to content

Guided tour

FabrCore 2.0 · Release and package availability

These guides track the current 2.0 source. Stable 2.0.0 publication is pending; package commands show the release target. Until it is published, follow the source quick start or use a matching available prerelease set. Release migration · Runtime modes

MODULE 09 · LESSON 9.4

Ingest policy documents into GraphRAG

Ingest policy documents with scope, provenance and visible job outcomes.

Lesson 51 of 86 · FabrCore 2.0

Overview

GraphRAG extraction creates searchable evidence and graph relationships. It needs compatible SQL schema and embeddings; in the current integrated implementation, GraphRAG expects 1536 dimensions. Ingestion instructions and model choice affect both cost and result quality.

Ingestion builds searchable evidence

GraphRAG extracts document evidence and relationships, then stores representations used for retrieval. It is more than uploading a file: parsing, model extraction and embedding generation can each fail or consume resources. Begin with a short fictional policy so you can compare the resulting evidence with every source paragraph before processing a large corpus.

Schema, models and scope must agree

The integrated implementation expects compatible SQL graph/vector objects and 1536-dimensional embeddings. The extraction model and instructions affect what facts and relationships are produced. Bind the document to the intended knowledge scope and retain its source identity so later answers can cite it and updates can be traced to the right source.

Ingest a small policy document

  1. Configure default/extraction and embeddings models, SQL feature storage and the intended knowledge scope.
  2. Submit the policy document through the supported ingestion service/API, preserving document identity, metadata and source. Inspect job progress and errors.
  3. Tune legacy top-level GraphRag:Ingestion settings only after measuring a baseline. Test repeat ingestion, cancellation and deletion before relying on the corpus.
ingestion-and-evals.md · reference snippet 
dotnet run --project src/FabrCore.Services.GraphRag.EvalConsole -- run --model graphrag --manifest src/FabrCore.Services.GraphRag.EvalConsole/corpus-policy-v2.json --max-chars 0 --response schema --sections-per-batch 8 --document-concurrency 2 --taxonomy-names current --relations current --endpoint-aliases off --result-cache off --taxonomy-cache off --embedding-cache off --reingest fresh --iterations 1

Inspect ingestion before asking questions

  1. Create a short policy document with a distinctive rule, such as “Escalate after two failed restart attempts,” and ingest it into a disposable scope using the illustrated worker/API path.
  2. Inspect the ingestion outcome and available metrics/errors. Verify searchable evidence contains that rule and a source reference. A successful upload response is not proof extraction completed.
  3. Ask for the rule through scoped retrieval, then test an unavailable/incompatible embedding configuration in staging. Expect an explicit readiness/ingestion failure instead of an empty index presented as a successful import.

This verifies the document-to-evidence pipeline. Large-corpus throughput and extraction quality require measurements with representative documents, not extrapolation from one file.

If the result is different

A model producing vectors of another dimension cannot be fixed by changing a version marker. Inspect schema, embedding metadata and extraction failures separately.

Go deeper

Explore the related documentation.