[johnnyhuy]

Allow the first document to be indexed into an initialised, empty Chroma collection, and let searches on that collection return no matches. Check explicitly for a missing adapter instead of using Chroma's collection length as an initialisation check.

Part of #3
Closes #26

The existing child issue #26 defines this slice. The parent remains open. Five new regression cases use temporary local Chroma storage and fixed embeddings, covering empty searches with and without a channel, the first insert and retrieval, and missing adapters. The existing retrieval fixture now inserts through Wingman's public wrapper too.

Verification

All commands ran with a cleared environment, MISE_NO_ENV=1 and UV_PYTHON=3.13. After installation, UV_FROZEN=1 kept the committed lockfile fixed. No real Slack or model tokens or .env files were read or changed.

• mise run install passed after trusting this new worktree's mise config. Installation refreshed unrelated lockfile wheel metadata, which was reverted before verification and commit.
• mise run test-backend passed with 77 tests. Before the fix, the same suite produced 3 failures, 69 passes and 5 fixture errors, all failures and errors from the empty collection guard.
• mise run test-frontend passed its existing placeholder command, No frontend tests yet. This is not frontend test coverage.
• mise run test-runtime passed all 7 tests with fixtures and local listeners.
• mise run dev-backend started on 127.0.0.1:18026, served the requests below with HTTP 200 and shut down cleanly after capture.
• git diff --check passed. No migrations or schema changes.

tests/test_retrieval.py::test_search_empty_collection_returns_no_matches[None] PASSED
tests/test_retrieval.py::test_search_empty_collection_returns_no_matches[CTARGET] PASSED
tests/test_retrieval.py::test_first_document_can_be_indexed_and_retrieved PASSED
tests/test_retrieval.py::test_missing_adapter_still_raises[add] PASSED
tests/test_retrieval.py::test_missing_adapter_still_raises[search] PASSED
======================= 77 passed, 112 warnings in 2.65s =======================
runtime tests 7, pass 7, fail 0

Offline proof

Backend and retrieval capture succeeded. The local backend used temporary SQLite and Chroma storage with fixed embeddings. A temporary startup harness disabled settings file loading, mocked Slack authentication, blocked Slack API calls and Socket Mode startup, blocked external network connections, and replaced runtime generation with a fixture response. It used the existing module-level vector_store and rag_engine singletons. The answer text below is a fixture response, not model output.

The before-and-after transcript ran the adapter from origin/main and this branch against an empty local collection with query and document embeddings fixed at [0.0, 1.0]. Scores are squared L2 distances, where lower is nearer; they prove the first document reached the index, not retrieval quality on real support questions. A missing score means no chunk was retrieved.

{"query":"How do I connect?","fixture_embeddings":[0.0,1.0]}
{"step":"empty_search_before_fix","error":"Vector store not initialized","chunks":[],"score":null}
{"step":"first_insert_before_fix","error":"Vector store not initialized","chunks":[],"score":null}
{"step":"empty_search_after_fix","chunks":[],"score":null}
{"step":"first_insert_after_fix","chunks":[{"content":"Use the IT portal to connect.","metadata":{"source":"api"},"squared_l2_distance":0.0}]}

The curl capture then exercised a separate fresh collection through the backend, using the default chunk settings.

curl --fail --silent --show-error -w '\nHTTP %{http_code}\n' http://127.0.0.1:18026/health
# {"status":"healthy"}
# HTTP 200

curl --fail --silent --show-error -w '\nHTTP %{http_code}\n' http://127.0.0.1:18026/api/ask \
  -H 'Content-Type: application/json' \
  -d '{"question":"How do I connect?","request_id":"fixture-empty"}'
# {"answer":"Fixture response only; no model call.","sources":[],"confidence":"low"}
# HTTP 200

curl --fail --silent --show-error -w '\nHTTP %{http_code}\n' http://127.0.0.1:18026/api/documents \
  -H 'Content-Type: application/json' \
  -d '{"title":"VPN guide","content":"Use the IT portal to connect.","source":"api"}'
# {"id":1,"message":"Document added successfully"}
# HTTP 200

curl --fail --silent --show-error -w '\nHTTP %{http_code}\n' http://127.0.0.1:18026/api/ask \
  -H 'Content-Type: application/json' \
  -d '{"question":"How do I connect?","request_id":"fixture-first-document"}'
# {"answer":"Fixture response only; no model call.","sources":[{"chunk":0,"source":"api","title":"VPN guide"}],"confidence":"low"}
# HTTP 200

The runtime fixture captured the following retrieved context after the insert. Before the insert, both context and chunks were empty.

{
  "query": "How do I connect?",
  "context": "From api:\nUse the IT portal to connect.",
  "chunks": [{
    "content": "Use the IT portal to connect.",
    "metadata": {"chunk": 0, "source": "api", "title": "VPN guide"},
    "squared_l2_distance": 0.0
  }]
}

No capture path was blocked. No dashboard UI changed, so a screenshot was not applicable. Live Slack delivery and external embedding/model services remain unverified by design. No Slack messages, DMs or live Socket Mode connections were made. Local proof files are not included in the code commit; this PR body records the transcripts.

GitHub CI passed lint-and-test and all three image builds at e0006b5. CodeQL and both Socket Security checks also passed.