Raw data

Everything the run produced, at the level it was produced. Aggregates on the rest of the site are computed from these files, not stored, so any figure here can be recomputed — or shown to be wrong.

What is in here

Per-document ingestion outcomes including failures, per-fact parsing results, per-query retrieval results, resource samples taken every five seconds during ingestion, and the corpus manifest listing every source URL. 11 documents, 29 questions.

Summary files

File	Contents
summary.csv	The comparison table, one row per platform
summary.json	The same figures with corpus metadata attached
manifest.json	Every corpus document, its source URL and its planted facts
manifest.full.json	The corpus before the size cap was applied
queries.json	Every question and the exact string that answers it

Per-platform results

Platform	Deployment	Run
fastgpt	fastgpt-deploy.json	fastgpt-run.json
dify	dify-deploy.json	dify-run.json
ragflow	ragflow-deploy.json	ragflow-run.json
anythingllm	anythingllm-deploy.json	anythingllm-run.json
openwebui	openwebui-deploy.json	openwebui-run.json
flowise	flowise-deploy.json	flowise-run.json

Reading the run files

documents[] — one row per corpus document, with ingested, the chunk count, elapsed seconds, and the error if it failed. Failures are present rather than filtered out.
parsing[] — per document, which planted facts were found and which were missed.
retrieval[] — per question, whether every accepted fact appeared in the retrieved context.
resource_samples[] — memory and CPU across the stack, sampled every five seconds for the length of the run.
chunks_are_exact — false when the platform exposes no chunk-listing API and parsing had to be approximated by retrieval.

Rerunning it yourself

The harness is on GitHub under MIT. It rebuilds the corpus from the same public endpoints, so a fresh run on your own hardware is comparable to this one. If your numbers differ from ours, we would like the issue.