Graph-RAG Retrieval & Answer Evaluation

Use connected records to prototype graph retrieval and assess evidence-grounded responses.

The record concepts include household composition, membership changes, household lifecycle, business lifecycle, ownership / founder context, workforce history and employment spells. Inspect retrieved relationships across the linked domains. Use the same questions and permitted source corpus. Separate supported claims from insufficient evidence.

This dataset is prepared to order. Choose the city, date range and required scope; the accepted order defines the population, final fields, formats and delivery.

Onboarded city options

Nairobi (Kenya) · Lagos (Nigeria)

Custom city requests

African Major Cities · Additional cities require availability, scope and quote confirmation.

Data type

Synthetic

Connected domains

Households, Business, Employment, Finance

Catalogue structure

Graph-connected

Key features

Record concepts for your order-specific extract.

Who it is for

• RAG engineers
• graph-AI teams
• evaluation researchers
• applied AI labs

Proposed study unit

Question with a dated graph / corpus snapshot

Explore graph retrieval and grounded answers against a matched non-graph comparison.

Record concepts for your extract

• household composition
• membership changes
• household lifecycle
• business lifecycle
• ownership / founder context
• workforce history
• employment spells

The agreed specification defines the final columns, keys, relationships and observation period for your order.

Hold the corpus, questions and language model constant; compare retrieval coverage, grounded answer accuracy and query cost using reviewed evidence paths.

Proposed questions for your custom order

1. Does graph-based retrieval improve evidence coverage and answer accuracy over a matched non-graph baseline for questions spanning households, businesses, employment and finance?
2. Does graph retrieval recover the relationships needed to answer a question?
3. Can the answer trace its claims to source nodes and records?
4. How do multi-hop answers compare with a flat retrieval baseline?
5. Does the system explain an absent relationship instead of inventing a link?

Proposed research and analytical applications

• Task prototyping
• Evidence tracing
• Error analysis
• Controlled evaluation

Study and modeling guidance

Information available at prediction or evaluation time

Retrieval may use only source records and relationships valid and available at the question's as-of date.

Comparison and evaluation

Hold the corpus, questions, model and generation settings constant. Use held-out cases, reviewed reference answers and source-ID scoring.

Incomplete observation

Treat missing or unverified evidence as insufficient evidence; absence of a relationship is not proof of a negative fact.

Dated relationships

Preserve stable source-record IDs and effective dates on evidence paths.

Dated relationships

Keep evaluation questions and reference answers out of training or prompt-tuning material.

Data generation

FCTE synthetic city engine

Data classification

Synthetic

FCTE generation methods and published city/domain checks provide source context. The order defines the final extract and the release-specific checks, documentation and limitations supplied with it.

Known limitations

• Reference answers and evidence paths require review before benchmark use.
• No executable baseline, reviewed answer set or benchmark scores are verified by the illustrative preview.
• Synthetic graph associations do not establish real-world relationships.
• Synthetic records describe generated scenarios; they do not establish observed population prevalence or real-world causal effects.

Catalogue base price

$695 USD

Price basis

Base price per selected city

Availability

Prepared after order

The displayed catalogue amount is a base price per selected city. Your requested dates, population, fields, formats and usage rights determine the agreed scope and final quote. Applicable published discounts and taxes are confirmed with that quote.

Your prepared package

Your order confirms the data files, documentation, evidence, examples and release controls included in the prepared package.

• Agreed data extract
• Schema and field documentation
• Release-specific evidence and limitations
• Agreed loading or analysis examples
• Licence, manifest, checksums and delivery note

SD_RESEARCH_DATASET/
├── data/
│   ├── graph_nodes.parquet
│   ├── graph_edges.parquet
│   └── source_records.parquet
├── tasks/
│   ├── task_definition.yaml
│   ├── questions.jsonl
│   ├── reference_answers.jsonl
│   └── evidence_paths.jsonl
├── evaluation/
│   ├── baseline_model.py
│   ├── metrics.json
│   └── example_notebook.ipynb
├── documentation/
│   ├── README.md
│   ├── DATA_DICTIONARY.csv
│   ├── SCHEMA.md
│   └── METHODOLOGY.md
├── evidence/
│   ├── CERTIFICATION.json
│   ├── VALIDATION_REPORT.pdf
│   └── KNOWN_LIMITATIONS.md
├── MANIFEST.json
├── CHECKSUMS.sha256
└── LICENSE.pdf

Delivery process

Confirm the cities, date range, records, formats, rights and price in writing. The order-specific extract is then prepared and supplied under the agreed payment and delivery conditions.

One secure download package for the agreed release and licensed entity.

Delivery timing, access expiry, formats and prerequisites are confirmed in the accepted order.

Support, refreshes, future releases, hosting and APIs are included only when expressly ordered.

Data Licensing Agreement · 2026-10

A non-exclusive, non-transferable licence for the named customer within the accepted scope and duration.

Permitted use

Internal research, development, testing and analytics within the accepted order.

Model training

Model training, external applications, commercial distribution and access by other legal entities must be expressly addressed in the order.

Redistribution

Raw or substantially reconstructable data cannot be resold, published or redistributed without separate written authorization.

Attribution

Preserve release provenance and applicable third-party notices. Published outputs must identify the data as synthetic.