Skip to content

2026-05-20-shyftr-phase-e-production-retrieval-kickoff.md

Latest commit

 

History

History
137 lines (98 loc) · 4.67 KB

2026-05-20-shyftr-phase-e-production-retrieval-kickoff.md

File metadata and controls

137 lines (98 loc) · 4.67 KB
Raw

ShyftR Phase E production retrieval kickoff

Date: 2026-05-20 AEST Repo: /Users/stefan/ShyftR Branch: main Starting HEAD: 6e15eef82922a86216439b1dfb1fdd4b69df4a1e Remote status: origin/main matches starting HEAD. Status: Phase E is ready to start from a clean pushed repository.

Starting truth

Phase D is closed and pushed in commit 6e15eef.

Verified before commit/push:

PYTHONPATH=.:src python -m compileall -q src scripts examples
python scripts/terminology_inventory.py --fail-on-public-stale
python scripts/terminology_inventory.py --fail-on-capitalized-prose
python scripts/public_readiness_check.py
git diff --check
git diff --cached --check
MEM0_TELEMETRY=False PYTHONPATH=.:src pytest -q \
  tests/test_benchmark_mem0_backend.py \
  tests/test_benchmark_adapter_contract.py \
  tests/test_benchmark_simple_baseline.py \
  tests/test_benchmark_longmemeval_standard_mapping.py \
  tests/test_benchmark_locomo_standard_mapping.py \
  tests/test_benchmark_locomo_mini_fixture.py \
  tests/test_benchmark_beam_standard_mapping.py \
  tests/test_benchmark_phase13_runner_controls.py

Observed result:

  • public readiness: PASS
  • focused benchmark regression: 47 passed in 3.12s
  • git diff --check: clean
  • pushed: main -> origin/main at 6e15eef

A supplemental swarm review lane (swarm2) reviewed public-safety risk before commit. It blocked committing generated benchmark JSON/HTML reports because those reports contained local paths and dataset/model-output text. The final commit therefore excluded reports/benchmarks/phaseD-*.json, reports/benchmarks/phaseD-benchmark-report.html, and docs/benchmarks/phaseD-benchmark-report.html.

Preserved but not committed

Phase C / native context-compaction and related future-phase WIP was not committed as part of Phase D cleanup because it did not pass the current staged gate and belongs to a different tranche boundary.

Preserved local copy:

/Users/stefan/shyftr-preserved-phase-c-g-wip-20260520

There is also a stash created during cleanup:

stash@{0}: WIP on main: e5a8a68 feat: complete phase b contract and readiness tranche

Do not apply the stash blindly before Phase E. Inspect it first if needed.

Phase E objective

Phase E: production retrieval substrate.

Goal: close the measured retrieval-quality gap with modern memory systems without sacrificing local-first auditability, review-gated doctrine, or privacy boundaries.

Phase E must be measured against Phase D baselines after material retrieval changes land. No public superiority claim is allowed from Phase E implementation alone.

Exact first tranche: E0 embedding provider contract

Start with E0, not E1/E2/E4.

Objective: separate test embeddings from real embeddings and make report metadata honest.

Likely files:

  • src/shyftr/retrieval/embeddings.py
  • src/shyftr/retrieval/vector.py
  • src/shyftr/benchmarks/runner.py
  • tests/test_embedding_provider_contract.py

Minimum RED tests:

  1. Default deterministic provider declares quality='test' or equivalent metadata.
  2. Real provider requires explicit selection and cannot be implied by ambient credentials.
  3. Provider metadata appears in pack/debug and benchmark reports.
  4. Dimension mismatch fails with a clear error.
  5. Redacted/private text is not sent to remote embedding providers unless explicit policy allows it.

Implementation steps:

  1. Define provider interface with embed(texts), metadata(), and privacy_mode.
  2. Keep deterministic default for tests.
  3. Add local model/provider integration only after explicit dependency choice.
  4. Document provider trust and privacy rules.

Focused verification for E0:

PYTHONPATH=.:src pytest -q tests/test_embedding_provider_contract.py tests/test_retrieval_vector.py

Broader Phase E guard after E0 is green:

PYTHONPATH=.:src python -m compileall -q src scripts examples
python scripts/terminology_inventory.py --fail-on-public-stale
python scripts/terminology_inventory.py --fail-on-capitalized-prose
python scripts/public_readiness_check.py
git diff --check

Non-goals for E0

  • Do not add hosted/cloud embedding defaults.
  • Do not infer real provider use from ambient credentials.
  • Do not send private/redacted text to a remote provider without explicit policy.
  • Do not add sqlite-vec/LanceDB/Qdrant behavior before the provider contract is pinned.
  • Do not claim benchmark improvement until a post-E Phase D rerun is performed.

Clean-start check before implementation

Before touching E0 code, run:

cd /Users/stefan/ShyftR
git status --short --branch
git rev-parse HEAD origin/main

Expected:

## main...origin/main
6e15eef82922a86216439b1dfb1fdd4b69df4a1e
6e15eef82922a86216439b1dfb1fdd4b69df4a1e