ThothII/.superpowers/sdd/2026-08-08-internal-qdrant-ollama/task-4-report.md

Task 4 Report — Narrow harness embedding configuration to internal Ollama

Status

Implemented on 2026-08-08 in /Users/mp/projects/ThothII/.worktrees/git-workspace-registry.

RED evidence

Command:

cd harness
./.venv/bin/pytest tests/test_internal_embeddings.py tests/test_config_resources.py -q

Observed before implementation:

• exit code 1
• 10 failed, 10 passed
• failures proved the missing OllamaInternalEmbeddings client and missing internal-only config validation

Representative failures:

• ImportError: cannot import name 'OllamaInternalEmbeddings'
• AttributeError: 'EmbeddingsConfig' object has no attribute 'provider'
• config tests DID NOT RAISE ConfigError for external provider, API key, and non-private base URL

GREEN evidence

Focused behavior suite:

cd harness
./.venv/bin/pytest tests/test_internal_embeddings.py tests/test_config_resources.py -q

• exit code 0
• 20 passed

Relevant harness verification:

cd harness
./.venv/bin/pytest tests/test_internal_embeddings.py tests/test_config_resources.py tests/test_ollama_ensure.py -q

• exit code 0
• 36 passed, 2 warnings

Changed-file lint:

cd harness
./.venv/bin/ruff check tht/config.py tht/config_compat.py tht/vectorstore/embeddings.py tht/cli/ollama_cmd.py tests/test_config_resources.py tests/test_internal_embeddings.py

• exit code 0
• All checks passed!

Patch hygiene:

git diff --check

• exit code 0

What changed

• translated schema-v3 resources.embeddings into the harness-compatible embedding config view
• validated the internal embedding contract only for that runtime-owned resources.embeddings path:
  • provider must be ollama_internal
  • model must be qwen3-embedding:0.6b
  • dimensions must be 1024
  • base URL must be http://embedding:11434 or loopback HTTP on port 11434
  • extra fields like api_key are rejected
• replaced the active embed client with OllamaInternalEmbeddings, using one bounded /api/embed request per batch
• removed task/query prefix rewriting from the active embedding path
• validated response count, vector dimension, and finite numeric values before returning embeddings
• kept tht ollama ensure --json stdout pristine while warming through the internal client

Self-review

• kept changes inside the brief-listed files
• preserved DWH and session-persistence behavior
• preserved the legacy OllamaEmbeddings import path as an alias to avoid unrelated call-site churn

Concerns

• the focused harness verification still emits two pre-existing warnings:
  • DeprecationWarning from testcontainers.postgres
  • FutureWarning because resources currently flows through the legacy config translation path

Fix round 1 — 2026-08-08

Findings addressed

• HIGH: external top-level embeddings remained an operational fallback and could still load
• MEDIUM: non-object embed JSON payloads escaped as raw AttributeError

RED evidence

Command:

cd harness
./.venv/bin/pytest tests/test_internal_embeddings.py tests/test_config_resources.py tests/test_ollama_ensure.py -q

Observed before the fix:

• exit code 1
• 2 failed, 36 passed, 2 warnings

Representative failures:

• AttributeError: 'list' object has no attribute 'get' from response.json() returning a JSON array
• Failed: DID NOT RAISE ConfigError for top-level external embeddings.provider=openai_compatible

GREEN evidence

Command:

cd harness
./.venv/bin/pytest tests/test_internal_embeddings.py tests/test_config_resources.py tests/test_ollama_ensure.py -q

Observed after the fix:

• exit code 0
• 38 passed, 2 warnings

Touched-file lint:

cd harness
./.venv/bin/ruff check tht/config.py tht/vectorstore/embeddings.py tests/test_internal_embeddings.py tests/test_config_resources.py

• exit code 0
• All checks passed!

Minimal fix

• validated the final active cfg.embeddings contract after config loading, so legacy top-level embedding inputs now fail explicitly unless they exactly match the internal Ollama contract
• converted non-mapping embed JSON payloads into controlled EmbeddingsError failures with sanitized diagnostics instead of raw attribute errors