feat(rag): overlapping chunker, retrieval knobs, sqlite-vec index

- Chunker: char overlap across boundaries + hard-split of oversized paragraphs.
- Retrieval knobs: rag_top_k / rag_min_score in settings + Settings UI.
- Vector index: sqlite-vec ANN over document embeddings, dual-written and
  backfilled, with brute-force cosine as the guaranteed fallback.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/test_documents.py

@@ -57,6 +57,46 @@ def test_search_empty_query_returns_nothing():
     assert asyncio.run(s.search_documents("", _fake_embed)) == []
 
 
+def test_chunker_overlap_and_hard_split():
+    s = _store()
+    # a single oversized paragraph (no blank lines, as in PDF text) is split
+    big = "x" * 2000
+    parts = s._chunk_text(big, size=800, overlap=120)
+    # each chunk is one <=size unit, plus at most an overlap tail (+separator)
+    assert len(parts) >= 3 and all(len(p) <= 800 + 120 + 2 for p in parts)
+    # consecutive chunks share an overlap tail
+    two = s._chunk_text("A" * 700 + "\n\n" + "B" * 700, size=800, overlap=120)
+    assert len(two) == 2 and two[1].startswith("A" * 120)
+
+
+def test_vec_index_used_and_matches_brute_force():
+    # On a host that can load sqlite-vec, the fast path must be exercised (not a
+    # silent fallback) and agree with brute force on the top hit.
+    s = _store()
+    if not s.vec_enabled:
+        import pytest
+        pytest.skip("sqlite-vec not loadable on this host")
+
+    async def run():
+        await s.add_document("Lego", "lego star wars boss fight tips", _fake_embed)
+        await s.add_document("GPU", "gpu vega vram notes", _fake_embed)
+        # vec table populated by the dual-write
+        conn = s._connect()
+        n = conn.execute("SELECT COUNT(*) FROM vec_documents").fetchone()[0]
+        conn.close()
+        assert n == 2
+
+        vec_hits = await s.search_documents("lego star wars", _fake_embed, limit=2, min_score=0.1)
+        assert vec_hits and vec_hits[0]["title"] == "Lego"
+
+        # force the brute-force path and compare the top title
+        s.vec_enabled = False
+        bf_hits = await s.search_documents("lego star wars", _fake_embed, limit=2, min_score=0.1)
+        assert bf_hits[0]["title"] == vec_hits[0]["title"]
+
+    asyncio.run(run())
+
+
 def test_extract_text_by_type():
     from synapse.main import _extract_text
     # plain text / markdown -> UTF-8 decode