2.8.1

Quilltap v2.8.1 Release Notes

Highlights

Built-in TF-IDF Embeddings - Semantic memory search now works out of the box with zero API keys required. The new built-in embedding provider uses TF-IDF with BM25 enhancement, running entirely offline.

First-Startup Seeding - Fresh installations now include a default character and embedding profile, making Quilltap immediately usable without configuration.

Ollama Quality-of-Life Improvements - Better defaults, auto-filled URLs, and fixes for common connection issues.

---

Major Features

Built-in TF-IDF Embedding Provider

• New qtap-plugin-builtin-embeddings provides zero-dependency, offline embedding
• Uses TF-IDF with BM25 enhancement and Porter stemming
• Bigram support for improved phrase matching
• Vocabulary automatically fits to your memory corpus
• Debounced refitting when memories change (5-second debounce)
• New database tables: tfidf_vocabularies, embedding_status
• Background jobs: EMBEDDING_GENERATE, EMBEDDING_REFIT, EMBEDDING_REINDEX_ALL
• API endpoints for manual refit and reindex operations
• UI displays vocabulary stats and embedding progress indicators
• Theme-aware provider badges via qt-badge-provider-* CSS classes

First-Startup Data Seeding

• Default character "Ben" seeded on first startup when database is empty
• Default "Built-in TF-IDF" embedding profile seeded and set as default
• Enables semantic memory search immediately without configuration
• Seeding runs as phase 1.25 after migrations in startup sequence
• Safe to call multiple times—only seeds when no data exists

---

Improvements

Connection Profile UX

• Auto-fill profile name with PROVIDER/MODEL_NAME when selecting a model
• Auto-fill base URL when selecting a provider with defaults (e.g., Ollama → http://localhost:11434)
• Connection test errors now display in the profile modal UI
• Improved error logging shows messages directly

Provider Fixes

• Ollama: Fixed 405 error caused by trailing slash in base URL creating double-slashes
• Ollama: Cheap LLM fallback now uses current profile's model instead of hardcoded llama3.2:3b
• Embedding providers: Fixed embedding-only providers not appearing in dropdown (plugin initialization now registers EMBEDDING_PROVIDER capability)

---

Bug Fixes

• Memory extraction: Handles LLMs returning object instead of string for content/summary fields
• Memory search: Regex metacharacters (*, ?, (), etc.) now escaped in fallback search, preventing crashes when semantic search unavailable