post-graph-rag

Graph RAG with a memory of time — on the PostgreSQL you already run.

post-graph-rag extracts entities and relations with an LLM, stores them as a property graph beside pgvector embeddings, and answers questions by fusing vector similarity, graph traversal and full-text search. What makes it different: a later document can close an earlier fact, so your model stops reporting that someone is both an ally and a rival.

No separate vector store. No graph engine to operate. One database, one consistency model, one backup — and transactions that span your graph and your application tables.

The graph layer underneath is post-graph, also usable on its own — more below.

⚡ Try it

pip install post-graph-rag

Point it at PostgreSQL with pgvector and any OpenAI-compatible endpoint:

from post_graph_rag import GraphRAG, RAGConfig

rag = GraphRAG(RAGConfig(
    api_base="https://your-router/v1", api_key=..., model="gemini-3.6-flash",
    embedding_model="gemini-embedding-001", embedding_dim=1536,
    db_uri="postgresql://localhost:5432/postgres", realm="my_kb"))
await rag.initialize()

await rag.index_document("Helena Voss is CFO of Calder Industries.", metadata={"source": "2019.txt"})
await rag.index_document("Priya Nair has been appointed CFO, succeeding Helena Voss.", metadata={"source": "2023.txt"})

answer = await rag.query("Who is the CFO of Calder Industries?")
# Priya Nair — the earlier fact is closed, not competing

→ Seven runnable examples · Full installation · Configuration reference

📈 Benchmarks

Paper: arXiv:2608.24921.

LongMemEval: long-horizon chat memory

On the full 500-question LongMemEval set — all six question types, nothing sampled — against the numbers Zep publish for Graphiti (arXiv:2501.13956):

Qualifications, so you can weigh them yourself: Zep judge with GPT-4o, this uses a three-model majority panel with the answering model excluded from its own jury; the generation models differ in cost class and vintage in a direction that cannot be signed; one question of 500 is excluded because neither extraction prompt could turn that session into triples. The harness, frozen config and every failing case ship in the repo — see the full write-up, which also documents the improvements that were tested and rejected.

ECT-QA: the adversarial corpus

Chat memory is the easy register. ECT-QA is the hard one — earnings call transcripts, sixteen quarters per company, every metric restated every quarter with only the date to tell the values apart. A store that merely accumulates cannot answer these at all; it has four values for "gross margin" and no way to choose.

Scored under the protocol ECT-QA's own authors use — an LLM judge comparing element-wise, with a refusal counted separately from a wrong answer:

A second judge from a different model family scores the same answers at 0.805 — two-tenths of a point apart, which matters more than either figure, since the usual objection to a judged rate is that it moves with the judge.

The breakdown is more useful than the total: incorrect elements sit at 0.14, refusals at 0.05, and relative-time questions produce no incorrect elements at all. When this system commits to a figure on this corpus, it is usually right.

Same qualifications apply, plus two specific to this comparison: their judge model is not available on our router, and their verbatim rubric is truncated in the public HTML, so ours reproduces their described categories rather than their text. Their figures are on their corpus slice; ours is 78 questions over 6 companies. Read a few points of margin as approximate rather than decisive.

🧱 The layer underneath: post-graph

A standalone library, if you want the graph without the RAG. It makes PostgreSQL behave like a graph database rather than emulating one:

• Table-per-vertex, table-per-edge — real foreign keys, real indexes, real constraints, so your graph is queryable by anything that speaks SQL
• Recursive CTE traversals — neighbours, paths and shortest-path with cycle detection, executed in the database rather than in your application
• Two levels of tenancy — realm for hard isolation (optionally schema-per-tenant), space for sub-grouping inside it
• pgvector on vertices and edges, searchable across live and historical rows
• Append-only history and trigger-based audit logging on every table, capturing old and new state with the acting user
• Promoted payload columns and server-side range queries — filter, order and bulk-delete on JSONB fields at the database, not in Python

pip install post-graph · Apache 2.0 · 635 tests

🌟 Why post-graph-rag?

Traditional Vector RAG systems suffer from "chunk isolation"—they retrieve isolated text passages based purely on semantic similarity, missing higher-level relationships and cross-document entity connections.

post-graph-rag solves this by building a dual representation inside PostgreSQL:

1. Unstructured Vector Passages: Full document chunks indexed with pgvector HNSW embeddings.
2. Knowledge Graph Triples: Extracted Subject-Predicate-Object entities connected by graph edges.
3. Structured Document Metadata: Rich metadata tracking (source, category, collection, document, page, paragraph, space).
4. Application-Level Space Sub-grouping (space): Scopes indexing and vector similarity search to application-specific environments (e.g. production, sandbox, staging, user_workspace) within {realm} tenant partitions.

Relationship quality

Extracted relations are context-specific, drawn from what the text actually states:

(Zeus) --[is_king_of]--> (Olympian gods)
(Zeus) --[son_of]--> (Cronus)
(Zeus) --[married_to]--> (Hera)
(Zeus) --[defeated]--> (Titans)

Vague connectors (relates_to, associated_with, connected_to, …), self-loops and blank endpoints are rejected at extraction time. A relation that reaches the graph always says something specific about the pair it connects, and two entities merely appearing near each other never produces an edge.

If the LLM cannot produce usable structure, indexing raises ExtractionError. Placeholder edges are never invented as a fallback: once written they are indistinguishable from genuine extracted structure.

Relations the text explicitly denies are stored with the positive predicate and negated: true, rather than as an inverted predicate like did_not_have_relationship_with. Traversal and synthesis can then exclude them instead of reading them as assertions.

Entity resolution

Entities are unique per (realm, space, lower(name)), enforced by a unique index, and are additionally resolved through aliases. Extraction records every other surface form it sees, so Babbage, Charles Babbage and C. Babbage converge on one vertex; the fuller name becomes canonical and the rest become aliases. Pronouns and relative references (he, his father, the company) are rejected outright — they cannot resolve to a stable vertex.

The same entity mentioned in many documents is one vertex, which is what allows the graph to connect chunks that share no vocabulary.

Chunking and document context

index_text() chunks a document (with overlap, so relations spanning a boundary survive) and threads a DocumentContext through the chunks — title, source, and the canonical entity names found so far. Without it, every chunk after the first is extracted blind and its pronouns become junk vertices.

Bring your own splitter by passing chunker= to GraphRAG, or use index_document() directly with your own DocumentContext.

rag = GraphRAG(config, chunker=my_splitter)
await rag.index_text(long_text, metadata=DocumentMetadata(document="babbage.txt"))

Community summarisation

Corpus-level questions — "what are the main themes here?" — cannot be answered by retrieving passages, because no single passage contains the answer. After indexing, cluster the entity graph and summarise each cluster:

await rag.index_text(doc_a, metadata=DocumentMetadata(document="a.txt"))
await rag.index_text(doc_b, metadata=DocumentMetadata(document="b.txt"))

await rag.build_communities()          # clusters + one LLM report per community

res = await rag.query("What are the main themes?", param=QueryParam(mode="global"))
print(res["retrieved_communities"])

Each report is stored as a vertex in communities with its own embedding, so global and hybrid retrieval find themes by similarity rather than by enumerating relations. Membership is recorded as community_members edges back to the entities, so a report is always traceable to the subgraph it came from.

Communities are derived data: build_communities() replaces the previous clustering for the space rather than accumulating stale clusters. Global mode degrades to relation ranking when none have been built, so it never hard-fails.

Detection uses Leiden (igraph + leidenalg, installed by default), falling back to deterministic label propagation if the native build is unavailable. Both are deterministic — a randomised partition would produce a different graph on every indexing run. Leiden is the default because partition balance matters: on the evaluation corpus the largest community holds 35% of the graph under label propagation against 17% under Leiden, and a community spanning a third of the graph summarises everything rather than a theme. Supply your own with community_detector=:

rag = GraphRAG(config, community_detector=my_detector)   # (nodes, edges) -> {node: community_id}

Repeated relations

The same triple extracted from several chunks is one edge whose weight increments, not several edges. Weight then breaks ties when ranking relations.

🧭 Exploration Support (1.10.0)

Three engine calls for exploration-first consumers — structure, coverage, change — with the agent loop staying yours:

# Structure: an opt-in topic tree above the flat communities
rag = GraphRAG(RAGConfig(community_levels=2))
tree = await rag.get_community_tree()

# Coverage: where has retrieval never looked? (opt-in telemetry, hash-only)
frontier = await rag.least_explored_communities(k=5)
dark = await rag.dark_entities(limit=100)

# Change: what moved since the last poll, from belief time
delta = await rag.changes_since(watermark)          # counts only, one round trip
if not delta.empty:
    detail = await rag.changes_since(watermark, summary=False)
watermark = delta.as_of

Hierarchy levels nest by construction (recursive supergraph clustering); level-filtered retrieval happens inside the vector search, not as a post-hoc trim; deltas are exactly-once under clock skew via database-clock watermarks; and re-indexing an unchanged document yields an empty delta. community_levels defaults to 1 — existing behaviour is untouched.

🏗️ Architecture Workflow

graph TD
    subgraph INDEXING ["1. Indexing"]
        A[Document + Metadata] --> B[Chunker]
        B --> C[Embedding Service]
        B --> D[LLM GraphExtractor<br/>validate · glean · normalise]
        C -->|chunk vectors| S[(PostgreSQL + pgvector<br/>via post-graph)]
        D -->|entities · triples · validity| S
        D -.->|later doc contradicts earlier| SUP[Supersession<br/>closes the old edge]
        SUP --> S
    end

    subgraph SCHEMA ["Tables in one database"]
        S --- V1[documents]
        S --- V2[entities]
        S --- V3[communities]
        S --- E1[relations<br/>valid_from/to · t_created/expired]
        S --- E2[doc_mentions]
        S --- E3[community_members / _children]
    end

    subgraph RETRIEVAL ["2. Retrieval — three channels, fused by RRF"]
        Q[Question] --> K[Keyword + subquery expansion]
        K --> C1[Entity vector search<br/>→ multi-hop traversal]
        K --> C2[Relation embedding search]
        K --> C3[Lexical / BM25 over relations]
        V2 --> C1
        E1 --> C1 & C2 & C3
        C1 & C2 & C3 --> F[RRF fusion<br/>MMR · node-distance rerank]
        V3 -->|global mode| F
        V1 -->|chunks| F
    end

    subgraph SYNTHESIS ["3. Synthesis"]
        F --> G[Temporal filter<br/>as_of · as_believed_at]
        G --> H[Prompt assembly<br/>renders each relation's validity]
        H --> I[LLM]
        I --> J[Answer + citations + triples]
    end

Three things in that diagram are the whole argument: supersession at write time, three retrieval channels fused rather than one, and validity rendered into the prompt — the last being worth +38 points on temporal reasoning on its own.

🧪 Runnable Examples

Seven scripts in examples/, each one capability, all runnable:

export OPENAI_API_KEY=...  OPENAI_API_BASE=...  POSTGRES_URI=...
cd examples && python 01_quickstart.py

📦 Installation

Install post-graph-rag via pip or uv:

pip install post-graph-rag

Or using uv:

uv add post-graph-rag

PostgreSQL Requirements

pgvector is required, not optional. Without it the vertex tables are created without embedding columns and every similarity search silently returns nothing. initialize() raises SchemaError if it is missing rather than degrading.

# macOS
brew install pgvector

# Debian/Ubuntu (match your server version)
sudo apt install postgresql-17-pgvector

CREATE EXTENSION IF NOT EXISTS vector;

🚀 Quick Start

1. Basic Indexing & Querying

import asyncio
from post_graph_rag import GraphRAG, RAGConfig, DocumentMetadata

async def main():
    # 1. Configure GraphRAG engine
    config = RAGConfig(
        api_base="http://localhost:4000/v1",       # OpenAI-compatible router endpoint
        api_key=os.environ["OPENAI_API_KEY"],     # Never hardcode credentials
        model="gemini-3.6-flash",                    # LLM model for extraction & synthesis
        embedding_model="gemini-embedding-001", # Embedding model
        embedding_dim=1536,                       # Must match the model's output width
        db_uri="postgresql://user:password@localhost:5432/postgres",
        realm="enterprise_kb",
        schema_per_realm=True                     # Give each tenant its own schema
    )

    rag = GraphRAG(config)
    
    # 2. Connect & initialize PostgreSQL graph schema
    await rag.initialize()

    # 3. Index unstructured documents
    doc_text = (
        "Zeus is the king of the Olympian gods, ruling sky and thunder from Mount Olympus. "
        "He is the son of Cronus and Rhea, and married to Hera. "
        "Zeus defeated the Titans in the Titanomachy to establish his rule."
    )
    
    result = await rag.index_document(doc_text, metadata={"source": "greek_mythology.txt"})
    print(f"Indexed document {result['document_id']}: Extracted {result['entities_extracted']} entities.")

    # 4. Perform Hybrid RAG Query
    response = await rag.query("Who are the parents of Zeus and what did he defeat?")
    
    print("\n=== SYNTHESIZED ANSWER ===")
    print(response["answer"])

    print("\n=== RETRIEVED GRAPH TRIPLES ===")
    for triple in response["retrieved_graph_triples"]:
        print(f"  - {triple}")

    # 5. Clean up
    await rag.close()

if __name__ == "__main__":
    asyncio.run(main())

📋 Document Metadata (DocumentMetadata)

post-graph-rag includes structured document metadata tracking via the DocumentMetadata model:

from post_graph_rag import DocumentMetadata

metadata = DocumentMetadata(
    source="https://mythology.org/zeus.html",  # Document origin (URL, filepath, API)
    category="greek_mythology",               # Document category/topic
    collection="olympian_deities",             # Collection namespace
    document="zeus_overview.pdf",              # Title or filename
    page=1,                                    # 1-based page number
    paragraph=2,                               # 1-based paragraph index
    space="production",                        # Sub-grouping within the realm
    extra={"author": "Homer", "year": -700}    # Custom metadata key-value pairs
)

await rag.index_document(chunk_text, metadata=metadata)

Design Rationale: Optional vs. Required

• All metadata fields are optional with default None. This allows seamless indexing of raw strings, short code snippets, webhooks, or unformatted text, while offering rich structural provenance tracking when indexing multi-page PDFs or categorized enterprise documents.

Document identity: source and document together

Re-indexing replaces rather than appends, so two documents that resolve to the same key are treated as one document seen twice — the second deletes the first.

The key is built from source and document together, so give at least one of them a value that is unique per document:

# Correct: source identifies this document
DocumentMetadata(source="/corpus/WDC-2022-q1.json", document="WDC-2022-q1")

# Wrong: a corpus name is not a document identity
DocumentMetadata(source="ect", document="WDC-2022-q1")   # was catastrophic before 1.8.0

Before 1.8.0 the key preferred source and ignored document entirely, so the second form collapsed an entire corpus onto one key. An 80-transcript run kept five transcripts and marked 92% of its relations dormant, with no error raised — the only visible symptom was the system declining to answer questions whose evidence had been deleted. Since 1.8.0 both parts contribute, so the second form is merely untidy rather than destructive.

Upgrading: keys computed before 1.8.0 do not match the new scheme, so re-indexing an existing document appends a copy instead of replacing it. Rebuild realms indexed on the old scheme.

⚙️ Configuration Reference (RAGConfig)

RAGConfig can be configured explicitly or automatically loaded from environment variables:

Environment variables are read when a RAGConfig is constructed, not at import time.

schema_per_realm

Off by default for backwards compatibility, but recommended for new deployments. With it off, every realm shares one physical set of tables filtered by a realm column — so the first realm to create entities fixes the embedding column width for all of them, and a second realm with a different embedding_dim cannot work.

allow_embedding_fallback

Off by default. Fallback vectors are not comparable with API embeddings, so mixing them into the same table corrupts retrieval rather than degrading it. With it off, an embedding failure raises EmbeddingError.

📖 API Reference

GraphRAG

The main orchestrator class for indexing and querying.

• await initialize(): Connects to PostgreSQL and creates the graph tables (documents, entities, relations, doc_mentions). Raises SchemaError if pgvector is unavailable or an existing table's embedding width disagrees with embedding_dim.
• await index_document(text, metadata=None, space=None) -> Dict[str, Any]: Embeds the chunk, extracts entities/triples via the LLM, resolves entities by name, and writes vertices, relations and doc_mentions edges. Raises rather than writing placeholder structure if extraction fails. Returns counts plus document_id and metadata.
• await query(question, param=None, top_k=None): Retrieves and synthesizes an answer. Returns question, answer, mode, keywords, retrieved_documents, retrieved_entities, retrieved_graph_triples, references. With QueryParam(stream=True) returns an async iterator of content chunks instead.
• await query_data(question, param=None) -> Dict[str, Any]: Structured retrieval with no synthesis — returns entities, relationships, chunks, references.
• await close(): Closes database connection pools.

QueryParam

• mode: one of mix, local, global, hybrid, naive, bypass. An unknown mode raises ValueError.
• top_k, max_total_tokens, max_entity_tokens, max_relation_tokens, response_type
  • The three token budgets default to None — unlimited — as of 1.11.1: everything retrieved reaches the model. Set an integer to cap context for cost or for a model with a small window.
• stream: return an async iterator of tokens instead of a dict.
• only_need_context: return retrieval output without calling the LLM.
• space: restrict retrieval to one space; __all__ queries across all spaces.
• conversation_history, hl_keywords, ll_keywords: supply keywords to skip extraction.

Errors

All inherit from RAGError, so failures surface instead of degrading into irrelevant results:

• SchemaError — pgvector missing, or embedding width mismatch.
• EmbeddingError — embedding request failed, or returned the wrong width.
• LLMError — completion or streaming call failed.
• ExtractionError — the LLM returned no usable entities or triples.

DocumentMetadata

Data container for structured document metadata.

• source: Optional[str]: Document URL, path, or origin.
• category: Optional[str]: Document category or domain.
• collection: Optional[str]: Document collection or folder.
• document: Optional[str]: File title or filename.
• page: Optional[int]: 1-based page number.
• paragraph: Optional[int]: 1-based paragraph index.
• space: Optional[str]: Sub-grouping space to index into.
• extra: Dict[str, Any]: Custom user metadata.
• to_dict() -> Dict[str, Any]: Serializes non-None fields to dictionary representation.
• from_dict(data: Dict[str, Any]) -> DocumentMetadata: Deserializes dictionary data.

RAGGraphStore

Database layer wrapping post-graph.

• add_document(text, embedding, metadata, space=None): Inserts a document vertex into the documents table.
• upsert_entity(name, entity_type, description, embedding, space=None): Upserts by canonical name within (realm, space), so an entity mentioned in many documents is a single vertex. A bare Concept stub never overwrites a richer type or description.
• find_entity_by_name(name, space=None): Resolve an entity vertex by name.
• add_relation(from_entity, to_entity, relation_type, description, space=None, embedding=None): Directed relation edge.
• add_doc_mention(doc_vertex, entity_vertex, space=None): Links a chunk to an entity it mentions.
• search_similar_entities(query_vec, top_k, space=None) / search_similar_documents(...): pgvector HNSW similarity search.
• search_similar_relations(query_vec, top_k, space=None): Semantic search over relation edges. Returns [] unless embed_relations is enabled.
• get_neighbors(entity_id, space=None): 1-hop outgoing relations, scoped to space.
• get_all_relations(limit, space=None): Relations with their endpoint vertices.

🗄️ PostgreSQL Database Schema

post-graph-rag automatically provisions and manages the following graph schema in PostgreSQL powered by post-graph:

🧪 Testing

pip install -e ".[test]"
createdb postgres && psql -d postgres -c "CREATE EXTENSION IF NOT EXISTS vector;"
POSTGRES_TEST_URI="postgresql://localhost:5432/postgres" pytest

DB-backed tests create a disposable schema per test realm and drop it afterwards. They skip automatically when PostgreSQL with pgvector is not reachable.

🤝 Contributing

Bug reports, failing test cases and pull requests are all welcome. CONTRIBUTING.md covers the parts that are specific to this project rather than generic advice: what the test suite needs, what it deliberately does not need, and the invariants worth understanding before changing them.

Issues tagged good first issue are real gaps rather than make-work — each one names the file to look at, what "done" means, and what you will learn from it.

Tests need PostgreSQL with pgvector and no LLM credentials — every test uses an in-process fake, so the suite is free, offline and deterministic.

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Developed by Chandan Rajah (chandan.rajah@gmail.com).