Update Anthropic model config for current Claude models

[hiryanchen]

Problem

backend/config.py pins ANTHROPIC_MODEL to claude-sonnet-4-20250514, which is deprecated and retires 2026-06-15. Upgrading off it isn't a one-line change, because two things in AIGenerator.base_params are incompatible with current models:

• temperature: 0 returns a 400. Current Claude models reject non-default sampling parameters. Bumping ANTHROPIC_MODEL on its own makes every query fail at API-call time — the server still starts and ingests documents fine (embeddings are local), so the failure only shows up once you ask a question.
• thinking is unset. Newer models run adaptive thinking by default, so part of the 800-token max_tokens budget goes to reasoning this app never reads. max_tokens caps thinking plus response text, so answers can get truncated.

Change

• config.py: pin ANTHROPIC_MODEL to claude-sonnet-5.
• ai_generator.py: drop temperature, add thinking: {"type": "disabled"} so the full max_tokens budget goes to the response.

Two files, no behavior change to the retrieval flow.

Note for maintainers

Both return sites in ai_generator.py (:90 and :138) do response.content[0].text unconditionally. That's safe with thinking disabled as set here, but it will raise if anyone enables thinking later — a thinking block can land at content[0]. Left alone to keep this PR scoped; happy to fix it here or in a follow-up if you'd prefer.

🤖 Generated with Claude Code