A boolean stream flag leaves callers guessing whether they get a string or a generator

The normalization item is the only one of the six that returns a wrong answer without raising anything ^[10]. Cosine similarity is the dot product of two vectors divided by their norms, so an inner-product index equals cosine only when both vectors already have length 1 ^[1]. FAISS's IndexFlatIP computes the inner product, and the post treats it as a cosine approximation that is valid only after you normalize ^[10]. The helper it ships divides by `np.clip(norms, 1e-10, None)` ^[11]. That clip matters: a zero-norm embedding divides by 1e-10 and comes out as zeros, so it scores zero against every query instead of turning the column into NaN ^[2].

The retry snippet is easy to check against its own numbers. `stop_after_attempt(4)` permits three sleeps, and `wait_exponential` doubling from the one-second minimum makes them 1, 2 and 4 seconds, or 7 seconds of waiting in total ^[12][3]. The `max=20` on the same line never engages inside four attempts ^[3]. `reraise=True` is the parameter that makes the next snippet work at all: with it, the final `RateLimitError` propagates unwrapped and the fallback's `except` clause catches it, while tenacity's default would wrap it in `RetryError` and the fallback would never run ^[5].

Both calls go through the decorated function ^[14]. A rate-limit event that hits the primary and the fallback therefore costs up to eight requests and about 14 seconds of sleep before the caller sees an error ^[4]. The post says pairing retries with a fallback is the same pattern most production LLM gateways use under the hood ^[15].

The history helper slices `self.history[-(self.max_history_turns * 2):]` ^[9] to avoid the orphaned assistant message that the post says some APIs reject outright ^[8]. Two times N is even, so the slice keeps pairs intact only while every entry in the list is exactly one user or one assistant message; a tool-result message in the same list puts the boundary back in the middle of a pair ^[6]. The cookbook these snippets come from includes a tool-calling template ^[16].

The first item is about a signature. Hiding streaming behind a flag means the call site cannot see whether it gets a string or a generator, and the post's complaint is that a wrong guess either fails silently or crashes deep in UI code ^[3]. Its fix is two names: `ask()` returns the full text, `ask_streaming()` yields chunks ^[4]. "Now the function signature is the documentation," the post says ^[2].

None of the six items is about the model's output ^[7]. Four of them are contracts in code the integrator writes. One is a numerical precondition and one is an availability policy ^[7]. The post does not report any measurements ^[8]. One claim does have to be checked locally before the fix is worth the refactor: whether an abandoned stream iteration leaks a socket depends on your SDK and its version, and the post says only that a naive implementation often does not close the connection ^[5].