v2.48.0

Release notes: v2.48.0

Memory & Tool RAG

• Content-hash skip for tool embeddings — tool_definitions stores embedding_input_hash; MemoryDB.upsert_tool() skips get_embedding() when the definition is unchanged and the hash matches. Reduces redundant API/local work on startup sync and manual sync_tools runs.
• ./bin/sync_tools.py cloud|local --force — Regenerates embeddings for every tool (ignores hash). Use after changing embedding model/dimensions or when debugging Tool RAG.
• Schema — New databases include the column; existing DBs add it via ALTER TABLE on open.
• Docs — docs/SYNC_ARCHITECTURE.md, docs/DUAL_DATABASE_SYSTEM.md (tool rows are not copied by sync-memory-db.py), docs/TOOL_RAG_STRATEGY.md, docs/EMBEDDING_HEALTH_CHECKS.md, bin/sync_tools.py usage header.

Embedding health check

• check-embeddings-health.py prints Embedding Provider as the effective vector backend (openai vs ollama), aligned with lib/embeddings.get_embedding() — not the chat LLM brand. When they differ (e.g. xAI chat + OpenAI embeddings), LLM Provider (chat) is shown on a second line.

Auto-memory & routing

• Auto-memory match hints and documentation; example AUTO_MEMORY_SIMILARITY_THRESHOLD 0.45 in docs.
• Conversation context and Tool RAG documentation updates; routing tweaks.

Tool RAG retrieval

• Typo hints — User-only token scan and segment matching to improve retrieval when queries don’t exactly match tool text.

Completion Guard

• Effective evidence, native-tool epochs, repair/tighten semantics.
• Yelp search tool and improved completion-guard context.

Search & media

• SerpAPI YouTube search and fallback tools; follow-up field title; embed YouTube in Web UI chat and Canvas.
• Canvas — Tooltip on page names in the gallery.

HTTP & tools

• LOCAL_PROXY2 proxy chain, tunnel retry, and proxy_used logging for HTTP clients using lib/http_client.py (and tools that rely on it).
• Per-bucket API rate limiting; phone_call timeouts hardened.

Spotify

• Honor device on play; episode lookup improvements; cleaner tool-retry speech.

Web UI

• Chat history title tooltips and full stored titles.

Voice & TTS

• TTS emoji cleanup and tool logging improvements.

Response style

• System prompt and response style doc updates; _format_auto_mode() keeps long complex-tool responses raw only when > 75 words (was a lower threshold).

Config & tests

• config/README updates.
• Schema tests assert tool_definitions.embedding_input_hash after fresh MemoryDB init (tests/test-db-schema-simple.py, tests/test-db-schema.sh).

---

Upgrade notes

• No manual migration required for the hash column; open any service or run sync_tools once to populate hashes.
• After switching embedding provider/model, run sync_tools.py with --force for the affected mode and re-check ./bin/check-embeddings-health.py.