Graph RAG Reporting

This project demonstrates how to build a powerful Graph RAG (Retrieval Augmented Generation) reporting application using the Perseus client. We'll show you how to convert unstructured PDF documents into a structured knowledge graph and leverage this graph to generate insightful, context-aware reports.

The Challenge: Extracting Actionable Insights from Complex Documents

Large corporate documents, such as annual reports and press releases, contain a wealth of information. However, extracting specific data points, understanding relationships between entities, and generating summarized reports can be a time-consuming and manual process. Traditional keyword-based search often lacks the semantic understanding required for deep analysis.

This leads to recurring issues for analysts and decision-makers:

• Information Overload: Sifting through hundreds of pages to find specific details.
• Lack of Context: Difficult to understand the "why" and "how" behind reported figures.
• Manual Reporting: Generating comprehensive reports requires significant manual effort and aggregation.

The Solution: Graph RAG with Perseus

This project leverages the Perseus Text-to-Graph engine to transform unstructured and semi-structured documents into a queryable knowledge graph. By combining knowledge graphs with Retrieval Augmented Generation (RAG), we can achieve more accurate, relevant, and context-aware report generation.

Why Graph RAG is crucial:

• Structured Knowledge: Information is extracted and stored as entities and relationships, making it easily queryable.
• Semantic Understanding: The graph preserves the context and relationships between data points, allowing for more nuanced retrieval.
• Enhanced Retrieval: RAG models can query the graph to retrieve highly relevant information, leading to more accurate generations.
• Dynamic Reporting: Generate reports on demand by asking natural language questions against the knowledge graph.

The Demonstration: L'Oréal Annual Report 2024

We process L'Oréal's 2024 Annual Report (PDF) to build a knowledge graph and then generate financial performance reports.

From Text to Structured Graph: An Example

Here’s how a section from a report is transformed into a structured entity within the knowledge graph, which can then be used for RAG.

📄 Source Text (LOREAL_Rapport_Annuel_2024.md):

L'Oréal s'appuie sur **37 marques internationales** réparties en 4 divisions :
**Produits Professionnels :** Kérastase, L'Oréal Professionnel, Redken, etc. Lancement du sèche-cheveu AirLight Pro.
...

🔍 Extracted Entity (in the Knowledge Graph):

type: BusinessDivision
label: Produits Professionnels
hasBrand -> Kérastase
hasBrand -> Matrix
hasBrand -> Redken
hasBrand -> L’Oréal Professionnel

This structured output forms the basis for rich, context-aware retrieval.

Powering the RAG: The simple-graph-retriever Package

Once the knowledge graph is created by the Perseus client and loaded into Neo4j, we need a way to efficiently search it to find context relevant to a user's query. This is the role of the simple-graph-retriever, a specialized package that bridges the graph database with the language model.

It performs two key operations in this workflow:

• Graph Indexing (index.py): The index.py script uses this package to process the graph in Neo4j. It applies the Leiden algorithm to identify coherent communities of related nodes within the graph. These communities are then broken down into smaller "chunks," which are embedded and indexed as vectors in the Qdrant database. This creates a searchable semantic layer on top of the graph.

• Graph Retrieval (report.py): The report.py script uses the package to take a natural language query (e.g., "Money KPIs"), search the Qdrant index for the most relevant chunks, and retrieve the corresponding subgraphs from Neo4j. This highly relevant, structured context is then passed to the LLM to generate an accurate report.

The Workflow at a Glance

git clone https://github.com/Lettria/perseus-client.git
cd perseus-client/examples/advanced/graph-rag-reporting

pip install -r requirements.txt
docker compose up -d

docker compose down

Generated Report Example

Here's an example of a report generated by querying the knowledge graph of the L'Oréal 2024 Annual Report with the query Money KPIs.

Visualize in the Perseus Interface

The knowledge graph outputs are accessible both locally (.ttl files) and in the console for interactive exploration:

• Nodes tab: Browse all extracted entities (e.g., companies, financial metrics, divisions).

• Graph tab: Visualize relationships interactively.

• Turtle and Cypher tab: View the Turtle (.ttl) and Cypher (.cql) files.