213 seconds per agent step evicts hybrid RAG from the local CPU

The weighted fusion step is worth reading slowly. Each engine's scores are min-max normalized, then blended at alpha equal to 0.7 ^[6]. Min-max is computed per query, so the best candidate in each list becomes 1.0 and the worst becomes 0.0 regardless of the raw numbers ^[15]. A query with no genuine lexical hit still hands its least-bad BM25 candidate a 1.0 before fusion ^[15]. The excerpt also never says which engine carries the 0.7; under a convex combination the other side gets 0.3, and which side that is decides whether this is a semantic ranker with lexical rescue or a keyword ranker with semantic smoothing ^[12].

The planning number is the step time. The author reports over 213 seconds per step on CPU, alongside Python sandbox import errors and hallucinated tool calls from the smaller local model ^[5]. Agent loops rarely finish in one step, and three steps at that rate is 639 seconds, close to eleven minutes for a single answer ^[13]. That arithmetic is what makes a hosted Qwen2.5-72B-Instruct look like the cheap option ^[4]. For the 213 seconds to mean anything on your hardware you would need what the excerpt does not supply: the CPU, the name of the local model, and the chunk count it was searching ^[17].

The index choice is good engineering and honestly labelled. IndexFlatIP over L2-normalized MiniLM vectors makes the inner product exactly cosine similarity ^[8], and the writeup is explicit that skipping quantization and clustering is what buys 100 percent accuracy in vector space ^[11]. Exactness comes from scanning everything, so per-query work rises with the number of vectors ^[14]. Chunks of 300 characters with 50 of overlap advance 250 characters at a time and duplicate about 17 percent of each chunk ^[16], which means the vector count climbs faster than the corpus does.

One tension in the writeup deserves a line of its own. It lists sandbox containment security flaws among the constraints to overcome ^[9] while also reporting Python sandbox import errors as a local blocker ^[5]. An import the sandbox refuses is the sandbox doing its job. If the resolution was widening what the agent's generated Python may import, that is a security decision rather than a troubleshooting note.

The writeup says it reports benchmark results ^[18]. The excerpt in front of me carries the architecture and the latency, not the retrieval scores, so the vocabulary-mismatch fix reads as a sound mechanism rather than a measured win ^[2][3]. The roadmap points at graphRAG and multi-agent consensus ^[10]. Consensus multiplies steps, and steps are precisely what the CPU measurement priced.