Comparing changes

Introduce the StorageBackend Protocol so the engine, ingest, browse, export, feedback, and MCP layers stop reaching into a vendor SDK directly. Two implementations live alongside the Protocol: - SQLiteBackend (default) — single file at ~/.context8/context8.db, three vec0 virtual tables for named dense vectors, FTS5 for native BM25 sparse search, JSON1 for tag arrays, WAL mode + 5s busy timeout. Zero infrastructure: stock pip install, no Docker, no daemon. - ActianBackend (legacy) — extracted from the old storage.py and preserved behind the new [actian] extra. Same external contract. StorageService is a thin facade that resolves the backend from CONTEXT8_BACKEND (default "sqlite") and delegates every Protocol method. Existing imports (`from context8.storage import StorageService`) keep working unchanged. Schema migrations are idempotent via apply_migrations(); a meta-table dim guard refuses to re-open a DB whose code_dim disagrees with the current CONTEXT8_USE_CODE_MODEL setting (vec0 dims are immutable). config.py adds CONTEXT8_BACKEND, CONTEXT8_DB_PATH, USE_CODE_MODEL. pyproject.toml adds sqlite-vec, moves Actian into [actian] (with allow-direct-references for the GitHub wheel URL), registers the "actian" pytest marker, and bumps to 0.5.0. tests/test_storage_sqlite.py covers schema migration idempotency, the dim guard, WAL pragma, vec0 round-trips, FTS5 special-character defanging, JSON1 tag filters, and scroll-cursor pagination (18 tests). Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

Drop every _require_actian()/av.* call from the search engine, browse, and export layers. They now talk to StorageService.search_dense(), search_sparse(), and scroll() — the same Protocol both backends implement, so the engine is fully backend-agnostic. - search/engine.py: rewritten around the new Protocol. find_duplicate, find_duplicate_or_variant, and search_by_solution all collapse to thin wrappers over storage.search_dense("problem"|"solution", ...). _build_filter now returns a SearchFilter dataclass; QueryAnalyzer weights and AttributionTracker integration are unchanged. - search/fusion.py (new): pure-Python Reciprocal Rank Fusion. Replaces av.reciprocal_rank_fusion. ~30 lines, k=60 by default, per-list weights for QueryAnalyzer-driven fusion tuning. - search/attribution.py: tolerant of both ScoredHit (record_id) and the legacy Actian point shape (id), so the tracker stays neutral. - browse.py / export.py: scroll() with a SearchFilter — JSON1 on SQLite, FilterBuilder on Actian. Cross-backend export/import round-trip verified: JSON has no vectors, import re-embeds. tests/test_e2e_sqlite.py mirrors the legacy Actian e2e against SQLite: collection shape, hybrid search, filtered search, named-vector access, ablation, feedback loop, attribution, quality boost, dedup, browse, export-import (24 tests, no infrastructure required). tests/test_search_filter.py exercises SearchFilter → SQL fragment + JSON1 translation as pure Python (10 tests). Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

start/stop print "no daemon needed" under SQLite (and only spin up the Actian container when CONTEXT8_BACKEND=actian). init runs schema migrations for SQLite or creates the Actian collection. doctor checks sqlite-vec, WAL mode, vec0 + FTS5 module presence, and runs a filtered-scroll probe — instead of poking the Actian client directly. stats shows "Database: SQLite (~/.context8/context8.db)" or the "Endpoint: host:50051" Actian line as appropriate. - cli/ui.py: check_backend() picks the probe by backend; check_db_connection is preserved as an alias. New check_sqlite_vec() reports the extension version. Old check_actian_sdk() kept for the doctor's Actian-specific section. - cli/commands/lifecycle.py: start/stop/init are dispatch + delegate. - cli/commands/serve.py: _bootstrap() skips Docker for SQLite, delegates to ensure_running() for Actian. --no-bootstrap flag unchanged. - cli/commands/ops.py: doctor splits into _doctor_sqlite() and _doctor_actian(). The old raw-client filter probe is gone — replaced with storage.scroll(SearchFilter(language="python"), 1) which works on both backends through the Protocol. browse/search/export/import all now route through check_backend() instead of the Actian-only check_db_connection() alias. - cli/commands/bench.py: "All Actian features + quality ranker" copy reworded to "Hybrid + filter + ranker (full pipeline)". demo's FilterBuilder mention now reads "Metadata filters (SQL WHERE on SQLite, FilterBuilder on Actian)". - docker.py: ensure_running, is_container_running, stop_container short-circuit to success when CONTEXT8_BACKEND != "actian", so legacy callers can't accidentally spin up Docker on a SQLite-default install. - mcp/tools.py: get_services() drops the unconditional Docker bootstrap — that already happens in serve._bootstrap. _handle_stats reports backend / db_path / endpoint as appropriate. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

README rewritten around the new SQLite-first install (single command: pip install context8 && context8 init --seed && context8 add claude-code). The Actian path is preserved as an "Optional: Actian VectorAI DB backend" section with the hackathon-era install. The "Hackathon: Advanced Features Used" section becomes "Capabilities (and how each backend delivers them)" — the same three capabilities (named vectors, hybrid fusion, filtered search) framed as backend-portable. Architecture diagram redrawn: pluggable Protocol fanning out to two concrete backends (SQLite vec0+FTS5 below, Actian gRPC container right). Tech-stack table promotes sqlite-vec and FTS5 to primary, demotes Actian to optional. Project-structure tree updated to reflect the new storage/ package and search/fusion.py. v0.5.0 changelog entry added. CLAUDE.md fully rewritten — new project overview, structure, key design decisions (#1 is now "pluggable storage backend"), commands, plus distinct SQLite Backend Notes / Actian Backend Notes sections. RESULTS.md kept as the hackathon submission narrative but reframed: the Actian-feature table is rewritten as a backend-portability table (SQLite delivers each capability via vec0/FTS5/SQL+JSON1; Actian delivers them via named vectors / sparse vectors / FilterBuilder). Benchmark section now has placeholders for both backends so the ablation can be run side-by-side. All twelve docs/*.md design docs (CONCEPT, ARCHITECTURE, BOTTLENECKS, PLAN-01..08, Hackathon Demo Video — Script) prepended with a historical-artifact banner pointing readers to README.md / CLAUDE.md. The hackathon design narrative is preserved in place; it just stops being the canonical source for the current architecture. tests/test_e2e.py: - pytestmark adds pytest.mark.actian + a new skipif on CONTEXT8_BACKEND != "actian", so the Actian e2e suite skips cleanly under the default SQLite install (15 tests skip). - Fixed pre-existing FeedbackService(storage, embeddings) arity mismatch on lines 281 and 300 — production constructor takes storage only (mcp/tools.py:44 confirms). Drop the second arg. - isolated_collection fixture now patches actian_backend.COLLECTION_NAME (the captured-at-import-time copy) rather than the package module attribute — the module's attribute is no longer load-time-bound to COLLECTION_NAME after the storage package split. Verification: 127 passed, 15 actian e2e skipped, ruff clean, context8 init/doctor/stats/search/bench/export/import all green under SQLite. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

`ruff format --check src/` is part of CI and was failing on nine files that were added or rewritten across the SQLite-backend rollout. This is purely whitespace and line-wrapping (no logic change) — the test suite is identical before and after. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

The version jump from 0.x to 1.0 marks two things: 1. The default install is finally a one-command `pip install context8` with no non-PyPI wheels. v0.x always required the Actian GitHub wheel; that wart is gone. 2. The public surface — MCP tools, CLI command names, env-var conventions, and the StorageBackend Protocol — is now committed to under semver. Future backend additions (e.g. Postgres + pgvector for a hosted version) will be additive third Protocol implementations, not breaking changes. Public surface guaranteed under semver going forward: - MCP tools: context8_search, context8_log, context8_rate, context8_search_solutions, context8_stats. - CLI commands: init, start, stop, doctor, stats, search, browse, add, remove, bench, demo, serve, export, import, import-github, mine. - Env vars: CONTEXT8_BACKEND, CONTEXT8_DB_PATH, CONTEXT8_DB_HOST, CONTEXT8_DB_PORT, CONTEXT8_USE_CODE_MODEL, CONTEXT8_TEXT_MODEL, CONTEXT8_CODE_MODEL, CONTEXT8_RECENCY_HALF_LIFE_DAYS. - Backend Protocol: StorageBackend + SearchFilter + ScoredHit in context8.storage. - JSON export format: {format: "context8-export", version: 1} (backend-agnostic, used for cross-backend migration). pyproject.toml: bump 0.5.0 → 1.0.0, classifier 3-Alpha → 4-Beta. src/context8/__init__.py: bump __version__. README: full v1.0.0 changelog entry rolling up everything since v0.4.0, plus a migration recipe for existing Actian users (export → upgrade → init → import re-embeds). Tag and push to trigger CI + PyPI publish: git tag v1.0.0 && git push --tags Verification: 127 passed, 15 actian e2e skipped, ruff clean, format clean, context8 --version reports 1.0.0. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

The v1.0.0 publish workflow failed with HTTP 400 from PyPI: Can't have direct dependency: actian-vectorai @ https://github.com/hackmamba-io/actian-vectorAI-db-beta/raw/main/ actian_vectorai-0.1.0b2-py3-none-any.whl ; extra == "actian" PyPI policy (PEP 440 §5) rejects packages that declare URL-pinned dependencies in their published metadata, even when gated behind an extra. The Actian SDK isn't on PyPI, so the only valid options are: (a) leave it out of pyproject.toml entirely and document the URL install, or (b) wait until it lands on PyPI. v0.4.0 used (a); we revert to that. Changes: - pyproject.toml: remove [project.optional-dependencies] actian and the [tool.hatch.metadata] allow-direct-references block. [tool.uv.sources] is preserved as a convenience for `uv pip install` from source. - README: install path becomes pip install context8 \ "actian-vectorai @ https://github.com/.../actian_vectorai-...whl" with a note explaining why it isn't a single command. - CLAUDE.md / RESULTS.md / config.py docstrings updated likewise. - src/context8/storage/actian_backend.py: ACTIAN_INSTALL_HINT no longer suggests the (now non-existent) [actian] extra; points at the wheel URL directly. Module docstring updated. - src/context8/storage/__init__.py: docstring updated. - src/context8/cli/ui.py: _check_actian and check_actian_sdk error messages point at the wheel URL. - tests/test_e2e.py: docstring install command updated. Verification: - `python -m build` succeeds; twine check PASSED on both .tar.gz and .whl. - Wheel METADATA inspected: zero `Requires-Dist:` URL pins (the actian-vectorai URL appears only inside the embedded README long-description, which PyPI doesn't validate). - ruff format --check / ruff check / pytest -q all clean (127 passed, 15 actian e2e skipped). Tag v1.0.0 will be moved to this commit and re-pushed; v1.0.0 never made it to PyPI so no consumers are affected. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Comparing changes

Open a pull request

Commits on May 2, 2026

This comparison is taking too long to generate.

Uh oh!