GraphBridge.GraphRag 0.2.0-preview.1

GraphBridge.GraphRag

Experimental. GraphRAG building blocks for GraphBridge, the EF Core provider for Neo4j. The API may change in any preview; using it reports GBRAG001, which you acknowledge once with <NoWarn>$(NoWarn);GBRAG001</NoWarn>.

Everything is stored through EF Core as typed entities, and every quote keeps the exact bytes it came from.

• Documents, sections and chunks:
  • EF entities linked child to parent (IN_DOCUMENT, SECTION_OF, IN_SECTION, NEXT_CHUNK, SUBSECTION_OF);
  • each chunk's text is exactly its document's UTF-8 bytes [StartByte, EndByte).
• Embeddings: a mapped float[] property, written by any Microsoft.Extensions.AI IEmbeddingGenerator.
• Resumable ingestion:
  • one transaction per document, and a lease per document;
  • unchanged documents are skipped;
  • transient model failures are retried, 4 times by default;
  • files named by a URI keep their documents' IDs wherever they are.
• Retrieval and answers:
  • vector and full-text search fused by rank;
  • an evidence budget and a trace of where each item came from;
  • answers whose every citation is located in the stored text, or a refusal with a reason;
  • search scoped to documents, sections, metadata or a Cypher query;
  • streamed answers, whose text is provisional until its citations are checked;
  • follow-up questions, rewritten with the conversation so far before they are retrieved;
  • answer export: one answer as one JSON document, whose citations Check confirms without the database;
  • derived text, such as OCR or a converted Office file: every citation names where the text came from;
  • experimental Text2Cypher, off by default: rows of a read-only query a model writes, behind a checked query and the server's plan.
• Ontologies and extraction: the standard generic document, novel and article ontologies, or your own classes.
  • Extraction prunes what the ontology doesn't allow and records why.
  • Every mention and quote is located in its chunk.
  • Things resolve by exact name; placeholders such as "the man" never merge.
  • Facts list every document that states them.
  • WithNeo4jGraphRagEntities() stores things in neo4j-graphrag's entity shape (__Entity__, FROM_CHUNK, name), so its retrievers read them.
• Maintenance: remove sources, in batches, reset what the toolkit wrote, clean up orphans, merge duplicate things with undo, and a health report.
• Agent tools: GraphRagTools gives any chat client or agent answer_from_documents, search_documents and, with an ontology, find_things and related_facts. Every result carries IDs and byte spans, and an optional trail records each call.
• Re-embedding: GraphRagReembedding.Plan(...) moves to a new embedding set through a staging property, with preview, backfill, cut-over and clean-up, never replacing vectors in place.
• Adopt mode: GraphRagAdoption.Plan(...) takes over a lexical graph Neo4j's Python neo4j-graphrag built, with a dry run, apply and undo; citations say whether they point into the source file or only into the stored chunk.
• Adapters: GraphBridge.GraphRag.DataIngestion for DataIngestion pipelines, GraphBridge.GraphRag.AgentFramework for Agent Framework agents, GraphBridge.GraphRag.AspNetCore for streaming endpoints, and GraphBridge.GraphRag.Pdf for PDF files with page-aware citations.
• Layouts: GraphRagLayout.Default, or GraphRagLayout.Neo4jGraphRag for the names Neo4j's Python neo4j-graphrag library reads.

// The model: dimensions must equal the embedding set's; the ontology is optional.
protected override void OnModelCreating(ModelBuilder modelBuilder)
    => modelBuilder.AddGraphRag(1536, o => o.Ontology = GraphRagOntologies.Article);

// The services: an IEmbeddingGenerator<string, Embedding<float>>, and an IChatClient for answers and extraction.
var set = new GraphRagEmbeddingSet("text-embedding-3-small", 1536);
services.AddGraphRag<AppDbContext>(set);

// The schema: constraints and indexes as typed migration steps, in your own reviewed plan.
var plan = new GraphMigrationPlan("app-002-graphrag", "app-001", GraphRagSchema.CreateSteps(GraphRagLayout.Default, set, GraphRagOntologies.Article),
    GraphMigrationRollbackPolicy.NoRollback("Dropping the indexes is a separate migration."));
await db.Database.ApplyGraphMigrationAsync(plan);

// Ingest the folder's Markdown, then ask, or query the graph with LINQ.
// .Folder reads *.md by default; pass "*.txt" for plain text.
var receipt = await ingestion.IngestAsync(GraphRagSources.Folder("articles"));
var answer = await answers.AnswerAsync("Who works for the library?");
var quotes = await db.Set<Quoted>().ToListAsync();

Samples 07 to 09 in the repository run offline. Requires Neo4j 2026.09 or later. Design: ADR 0052 in the repository's docs/architecture/adr.