Updated parse_azure_endpoint in utils.py:200 to strip the /openai/deployments/... path from the endpoint URL. Previously it only removed the query string, leaving the full deployment path — which AsyncAzureOpenAI then duplicated, causing the 404.

 Updated test_online.py to use create_chat_model() → model._model.request() — the same _make_azure_provider → AzureProvider(openai_client=AsyncAzureOpenAI(...)) code path used by the rest of the codebase.

fixes bugs described in #238 
- regression URL parsing in PR #231
- uv.lock updated to newer versions in many cases

**Stack: 3/4** — depends on #229. Merge #231, #229, then this PR.

---

- Add `add_terms_batch` and `add_properties_batch` to
`ITermToSemanticRefIndex` and `IPropertyToSemanticRefIndex` interfaces
- SQLite backend uses `executemany` instead of individual
`cursor.execute()` calls (~1000+ calls per indexing batch reduced to
2-3)
- Restructure `add_metadata_to_index_from_list` and
`add_to_property_index` to collect all data first (pure functions), then
batch-insert
- Memory backend implements batch methods as loops for interface
compatibility

## Benchmark

### Azure Standard_D2s_v5 -- 2 vCPU, 8 GiB RAM, Python 3.13

#### Indexing Pipeline (pytest-async-benchmark pedantic, 20 rounds, 3
warmup)

Only the hot path (`add_messages_with_indexing`) is timed -- DB
creation, storage init, and teardown are excluded.

| Benchmark | Before (min) | After (min) | Speedup |
|:---|---:|---:|---:|
| `add_messages_with_indexing` (200 msgs) | 28.8 ms | 25.0 ms |
**1.16x** |
| `add_messages_with_indexing` (50 msgs) | 7.8 ms | 6.7 ms | **1.16x** |
| VTT ingest (40 msgs) | 6.9 ms | 6.1 ms | **1.14x** |

Consistent ~14-16% improvement -- `executemany` amortizes per-call
overhead.

<details>
<summary><b>Reproduce the benchmark locally</b></summary>

Save the benchmark file below as
`tests/benchmarks/test_benchmark_indexing.py`, then:

```bash
pip install 'pytest-async-benchmark @ git+https://github.com/KRRT7/pytest-async-benchmark.git@feat/pedantic-mode' pytest-asyncio

# Run on main
git checkout main
python -m pytest tests/benchmarks/test_benchmark_indexing.py -v -s

# Run on this branch
git checkout perf/batch-inserts
python -m pytest tests/benchmarks/test_benchmark_indexing.py -v -s
```
</details>

---

*Generated by codeflash optimization agent*

---------

Co-authored-by: Bernhard Merkle <bernhard.merkle@gmail.com>

**Stack: 4/4** — depends on #230. Merge #231, #229, #230, then this PR.

---

- Five call sites used `get_item()` per scored ref — one SELECT and full
deserialization per match (N+1 pattern)
- Added `get_metadata_multiple` to `ISemanticRefCollection` that fetches
only `semref_id, range_json, knowledge_type` in a single batch query
- Replaced the N+1 loop with one `get_metadata_multiple` call at each
site
- Further optimized scope-filtering: binary search in `contains_range`,
inline tuple comparisons in `TextRange`, skip pydantic validation in
`get_metadata_multiple`

### Call sites optimized

1. `lookup_term_filtered` — batch metadata, filter by
knowledge_type/range
2. `lookup_property_in_property_index` — batch metadata, filter by range
scope
3. `SemanticRefAccumulator.group_matches_by_type` — batch metadata,
group by knowledge_type
4. `SemanticRefAccumulator.get_matches_in_scope` — batch metadata,
filter by range scope
5. `get_scored_semantic_refs_from_ordinals_iter` — two-phase: metadata
filter then batch fetch

### Additional optimizations

- **Binary search in `TextRangeCollection.contains_range`**: replaced
O(n) linear scan with `bisect_right` keyed on `start`, reducing
scope-filtering from ~25ms to ~9ms
- **Inline tuple comparisons in `TextRange`**: replaced `TextLocation`
allocations in `__eq__`/`__lt__`/`__contains__` with a shared
`_effective_end` returning tuples
- **Skip pydantic validation in `get_metadata_multiple`**: construct
`TextLocation`/`TextRange` directly from JSON instead of going through
`__pydantic_validator__`

## Benchmark

### Azure Standard_D2s_v5 — 2 vCPU, 8 GiB RAM, Python 3.13

#### Query (pytest-async-benchmark pedantic, 200 rounds)

200 matches against a 200-message indexed SQLite transcript. Only the
function under test is timed.

| Function | Before (median) | After (median) | Speedup |
|:---|---:|---:|---:|
| `lookup_term_filtered` | 2.650 ms | 1.184 ms | **2.24x** |
| `group_matches_by_type` | 2.428 ms | 978 μs | **2.48x** |
| `get_scored_semantic_refs_from_ordinals_iter` | 2.541 ms | 2.946 ms |
0.86x |
| `lookup_property_in_property_index` | 25.306 ms | 9.365 ms | **2.70x**
|
| `get_matches_in_scope` | 25.011 ms | 9.160 ms | **2.73x** |

<details>
<summary><b>Reproduce the benchmark locally</b></summary>

```bash
pip install 'pytest-async-benchmark @ git+https://github.com/KRRT7/pytest-async-benchmark.git@feat/pedantic-mode' pytest-asyncio
python -m pytest tests/benchmarks/test_benchmark_query.py -v -s
```
</details>

---

*Generated by codeflash optimization agent*

---------

Co-authored-by: Bernhard Merkle <bernhard.merkle@gmail.com>

fixed regressions which were caused by #231 that only showed when using real API keys.