mnemosyne-engine v1.1.0 -- document ingestion, partition refactor, sc…

…hema ingestion

mnemosyne-ollama v0.1.1 -- partition visibility, version flag fix

mnemosyne-mcp v0.2.0 -- federated search, search_docs, schema tools, …

…truncation fix

Add mnemosyne-ollama v0.1.0, bump engine to v1.0.5, fix PyPI docs

  New package: mnemosyne-ollama (ollama/)
  - Lightweight MCP host bridging Ollama to mnemosyne-mcp
  - Zero new dependencies (stdlib urllib, mcp SDK via mnemosyne-mcp)
  - Auto-detects tool-capable models (Gemma 3/4, Llama 3.x/4, Qwen, Phi-4)
  - Single-shot and interactive CLI modes
  - 32K context window for full search result delivery
  - OIDC publish workflow (publish-ollama.yml)

  Engine v1.0.5 fixes:
  - README: relative links converted to absolute GitHub URLs for PyPI rendering
  - REFERENCE: 10 Mermaid blocks replaced with pre-rendered PNGs
  - publish.yml: reject ollama-v tags to prevent cross-publish
  - Trademark notices added to README and ollama README

  Includes all pending engine, test, and config changes.

Add mnemosyne-ollama v0.1.0, bump engine to v1.0.5, fix PyPI docs

  New package: mnemosyne-ollama (ollama/)
  - Lightweight MCP host bridging Ollama to mnemosyne-mcp
  - Zero new dependencies (stdlib urllib, mcp SDK via mnemosyne-mcp)
  - Auto-detects tool-capable models (Gemma 3/4, Llama 3.x/4, Qwen, Phi-4)
  - Single-shot and interactive CLI modes
  - 32K context window for full search result delivery
  - OIDC publish workflow (publish-ollama.yml)

  Engine v1.0.5 fixes:
  - README: relative links converted to absolute GitHub URLs for PyPI rendering
  - REFERENCE: 10 Mermaid blocks replaced with pre-rendered PNGs
  - publish.yml: reject ollama-v tags to prevent cross-publish
  - Trademark notices added to README and ollama README

  Includes all pending engine, test, and config changes.

Bump to v1.0.4 — fix publish workflow SHA, update version refs

Bump version to 1.0.3 for trusted publisher verification

v1.0.2 — Fix self-ingestion bug, benchmark root path, README overhaul (…

…#21)

Self-ingestion fix (config.py):
- Added "mnemosyne" to default ignore_patterns so the engine does not
index its own source package when run from the repo root or when the
mnemosyne/ directory exists in a project.
- Removed manual workaround from benchmark_suite.py and benchmark.py
that was compensating for this missing default.

Benchmark root fix (httpx.json):
- Changed empty "root" field to "/tmp/mnemosyne_benchmark_httpx_0.28.1"
matching httpx_holdback.json. Empty root resolved to the mnemosyne
package directory, causing all 6 httpx questions to return 0% recall.
- Post-fix: httpx-6 file recall 0% → 80.6%, aggregate 76.8% → 87.3%.

Benchmark cleanup:
- Made single-project benchmark generic (removed hardcoded project
questions and compression targets, updated docstrings and report title).
- Removed product-specific references from TUNING.md.

README overhaul:
- Condensed README.md from 1,280 lines to ~120 lines. Centered logo,
standard GitHub project structure (install, benchmarks, features, use
cases including non-LLM scenarios), branding moved to footer.
- Created REFERENCE.md with full CLI docs, config, architecture, key
innovations, and integration guides.
- Replaced pipeline.png with mnemosyne.png as repo logo.
- Updated .gitignore exception accordingly.

Version bump: 1.0.0 → 1.0.2
- __init__.py, pyproject.toml, CHANGELOG.md, PLAYBOOK.md,
requirements.txt

[v1.0.0] Dense embeddings, retrieval quality upgrades, README update …

…for v1.0.0 (#19)

Summary

- Dense embedding backend — optional 23MB ONNX model
(all-MiniLM-L6-v2-code-search-512) as 6th retrieval signal.
  Opt-in via config, one-time local download, zero data egress.
- Porter stemmer in TF-IDF tokenizer — aligns with BM25/FTS5 porter
stemming
- Code-aware stopwords — preserves get, set, for, is, has in code
identifier splitting
- Two-pass soft file filter — chunks from top-50 individually scored
chunks survive at 0.7x penalty
- Symbol match file promotion — guaranteed filter slots for strong
symbol matches
- FTS5 escape fix — commas/periods/hyphens no longer break BM25 queries
- Type-aware symbol boost — 3x all, 4x for class definitions with
PascalCase/TitleCase detection
  - Decorator inclusion in function/class chunks
- Chunk enrichment for TF-IDF embeddings (file path + module docstring
context)
- README overhaul with benchmark comparison table, LLM integration
guide, pipeline diagram
  - Version bump to 1.0.0, Development Status → Production/Stable

  Benchmarks (Claude Opus 4.6, real codebases)

| ---------------------│ Production (829 files) │ Open-source (100
files)
│ Token cost │ 12K vs 45K (73% savings) │ ~4K vs ~3.5K
│ Correct file found │ 100% (20/20) │ 100%
│ Answer quality │ Equivalent to baseline │ Equivalent to baseline

  Resolves

  - Closes #5 (dense embeddings for semantic search)
  - Closes #6 (query expansion — replaced with stemmer + dense)
  - Closes #7 (symbol type-aware ranking)
  - Closes #9 (hit@3 refinement)
  - Closes #10 (health CLI + structured logging)

  Test plan

  - 306 non-benchmark tests pass
  - 5 benchmark gate tests pass
  - Dense features opt-in, zero impact when disabled
  - All version references updated to 1.0.0

Bump v0.4.1 — update README version references for PyPI (#17)

README.md still showed v0.3.0 in the version badge and throughout the
architecture docs. Updated all references to v0.4.0. Bumped package
version to 0.4.1 for PyPI re-publish (0.4.0 already uploaded).