Write the chat and query prompts in Russian #40

2026-07-31T15:48:59+02:00

claude commented

2026-07-31 15:48:59 +02:00

The chat and query system prompts were English, a Russian persona block gets
prepended in front of them at runtime, and the reply has to be Russian. Half
and half, on a 0.8B. This writes them in Russian.

Also deletes the You are maven, a self-hosted personal assistant opener. The
persona block one line above already says who she is, so that was a second
self-introduction with a lowercase name. The same duplicate came out of
KnowledgePrompt earlier.

The JSON contract line and the mood values are untouched — those are a wire
format, not prose.

Measured on the talk fixture, 27 cases, three runs each, Qwen3.5-0.8B at 4096:

	grammar only	+ Russian prompts
overall	14, 15, 8	11, 17, 15
chat	1, 3, 0	1, 5, 3
query	5, 6, 3	5, 6, 8
knowledge	8, 6, 5	5, 6, 4
p50 latency	~11.5s	3.0, 2.5, 2.2s

Chat and query went up, knowledge went down, and every one of those moves fits
inside the run-to-run spread. So the honest read is "not worse, probably better
on the two paths it targeted, not provable in three runs". Chat is the most
convincing because it started on the floor.

The latency drop is real and showed up in all three runs. Shorter prompts are
part of it; she also seems to have stopped writing English reasoning before the
answer, which is fewer tokens to generate.

What did NOT improve: address is still 21-22/27 and feminine slipped to
24-25. She writes вашей, подождите, делаете, я готов, я понял even
though the prompt now forbids exactly those. Telling a small model "never do X"
is not working — that is the argument for #402, and for templates if #402 fails.

The same run turned up two bugs that are not this PR fault, fixed separately:
the grammar truncates some replies to {, and the eval nonempty check passes
them.

Leftovers filed as #404: PhraseReminder is still fully English, and two prompts
still open with an English Respond ONLY with valid JSON:.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ

The chat and query system prompts were English, a Russian persona block gets prepended in front of them at runtime, and the reply has to be Russian. Half and half, on a 0.8B. This writes them in Russian. Also deletes the `You are maven, a self-hosted personal assistant` opener. The persona block one line above already says who she is, so that was a second self-introduction with a lowercase name. The same duplicate came out of KnowledgePrompt earlier. The JSON contract line and the mood values are untouched — those are a wire format, not prose. Measured on the talk fixture, 27 cases, three runs each, Qwen3.5-0.8B at 4096: | | grammar only | + Russian prompts | |---|---|---| | overall | 14, 15, 8 | 11, 17, 15 | | chat | 1, 3, 0 | 1, 5, 3 | | query | 5, 6, 3 | 5, 6, 8 | | knowledge | 8, 6, 5 | 5, 6, 4 | | p50 latency | ~11.5s | 3.0, 2.5, 2.2s | Chat and query went up, knowledge went down, and every one of those moves fits inside the run-to-run spread. So the honest read is "not worse, probably better on the two paths it targeted, not provable in three runs". Chat is the most convincing because it started on the floor. The latency drop is real and showed up in all three runs. Shorter prompts are part of it; she also seems to have stopped writing English reasoning before the answer, which is fewer tokens to generate. What did NOT improve: `address` is still 21-22/27 and `feminine` slipped to 24-25. She writes `вашей`, `подождите`, `делаете`, `я готов`, `я понял` even though the prompt now forbids exactly those. Telling a small model "never do X" is not working — that is the argument for #402, and for templates if #402 fails. The same run turned up two bugs that are not this PR fault, fixed separately: the grammar truncates some replies to `{`, and the eval nonempty check passes them. Leftovers filed as #404: PhraseReminder is still fully English, and two prompts still open with an English `Respond ONLY with valid JSON:`. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ

claude added 1 commit 2026-07-31 15:48:59 +02:00

Write the chat and notes prompts in Russian c7dadc97d9

The reply has to be Russian, but two of the phrasing prompts told her
what to do in English. Both are Russian now, in the same style as the
nudge prompt that already works better.

Also dropped the "you are maven, a self-hosted personal assistant"
line from both. The persona block right above it already says who she
is, so it was said twice.

The JSON part is unchanged.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ

kami commented

2026-07-31 20:22:37 +02:00

Superseded by #47, which landed this whole stack on master as one reviewed integration merge. This PR head is an ancestor of master — its commits are in, nothing here is lost. Closing as merged-by-proxy rather than merged, since the merge came in through #47.

Review threads on this PR were answered or acted on before the merge; the Russian wording fixes went in as #48.

Superseded by #47, which landed this whole stack on master as one reviewed integration merge. This PR head is an ancestor of master — its commits are in, nothing here is lost. Closing as merged-by-proxy rather than merged, since the merge came in through #47. Review threads on this PR were answered or acted on before the merge; the Russian wording fixes went in as #48.

kami closed this pull request

2026-07-31 20:22:38 +02:00

Pull request closed

Please reopen this pull request to perform a merge.

Sign in to join this conversation.

2 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: kami/Maven#40