TOOL VERDICTS

RAGFlow 1.0 Claims 100x the Concurrency

RAGFlow 1.0 shipped November 6, rewritten in Go, repositioned as an agent native database. The vendor's claim is 100x the concurrency and 90 percent token savings on a hit. The real question is whether the agent uses the compiled context at all.

2026-11-12

RAGFlow 1.0 shipped November 6, 2026, rewritten in Go and repositioned as an agent native database. The announcement is worth reading carefully because it states the account’s thesis back to us as a product: the center of gravity in retrieval is shifting from recall algorithms to orchestration frameworks plus context supply. The harness orchestrates. RAGFlow supplies.

The 36-second reel: RAGFlow 1.0 claims 100x the concurrency.

100x the concurrency

The Go rewrite is the headline number. After the rewrite, the vendor says retrieval concurrency improved by more than two orders of magnitude over the Python version, with container images about 10x smaller and startup about 10x faster. Treat the 100x as the vendor’s claim, not an independent measurement. What matters is the direction: retrieval is no longer a single function call that queries once and returns a few snippets. It is multi-turn orchestration, and the harness needs a layer that can keep up.

Compile once, benefit many times

Knowledge compilation is the core capability. It shifts the understanding work from query time to index time. Parsing, chunking, extraction, and compilation happen when the document lands, not when the agent asks. The compiled artifacts are claims with verbatim evidence, entity relations, timelines, and ontology labels, typed, versioned, and governable. The agent starts with facts and quotes instead of raw chunks. That is the whole pitch in one sentence.

The token economics

The vendor’s numbers, per the announcement: on an evidence-chain hit, a single query saves over 90 percent of tokens. On a miss, it costs about 2 percent. The break-even hit rate is about 2 percent, and on the FRAMES multi-hop benchmark the measured hit rate approached 43 percent with a free compilation model, saving 30 to 40 percent of tokens on average. The asymmetry is the point. Hits pay far more than misses cost, so compiling ahead beats reconstructing from scratch every query.

The test: quote or hallucinate?

Compiled context only helps if the harness uses it. This account has seen agents ignore good context before, so the reel ends with a test, not a verdict. Run your agent on compiled context and watch what it does with the evidence. Does it quote the source, or does it hallucinate? The answer tells you whether the supply layer is doing its job or whether the agent is still doing its own thing. That is the number to measure before adopting any of this.

This post started as an Instagram post →

Discussion

Talk it through

Argue with us on Instagram.