release: 0.2.5860

CHANGELOG.md

@@ -2,6 +2,11 @@
 
 ## Unreleased
 
+## 0.2.5860 - 2026-09-27
+
+- `codedb semantic-index` sends larger batches with more requests in flight, so hosted index builds finish faster.
+- Index builds wait out hosted rate limits with a longer capped backoff instead of failing after a few quick retries. Interactive `codedb context` calls still fail fast to local results.
+
 ## 0.2.5859 - 2026-09-27
 
 - Fix `codedb context` reporting `semantic lane: unavailable` with the default hosted configuration. The hosted default model id is now `codedb-hosted-v1`. Local ANN sidecars built by earlier releases are rebuilt once in the background; custom endpoints and model overrides are unchanged.

README.md

@@ -489,7 +489,7 @@ endpoints/model overrides and new repositories are never auto-built; set
 never mixes vectors from different models.
 
 It splits already-indexable files into bounded 832-byte source chunks, uses
-four concurrent 25-item requests by default (explicitly configurable from one
+eight concurrent 100-item requests by default (explicitly configurable from one
 to eight), and writes a
 small `semantic-chunks-v3.meta` mapping plus a generation-named `.hmls` mmap
 slab only in codedb's per-project local data directory. The directory and

build.zig.zon

@@ -1,7 +1,7 @@
 .{
     .name = .codedb2,
     .fingerprint = 0x6e18d96ca2a31757,
-    .version = "0.2.5859",
+    .version = "0.2.5860",
     .minimum_zig_version = "0.17.0-dev.813+2153f8143",
     .dependencies = .{
         .nanoregex = .{

docs/cli.md

@@ -100,7 +100,7 @@ command returns the local result and does not run an embedding model on CPU.
 default; `--local` (or `--no-semantic`) is the explicit on-device-only mode.
 
 `semantic-index` is the explicit ANN build trigger. It sends bounded 832-byte
-code chunks using four concurrent 25-item requests by default (configurable
+code chunks using eight concurrent 100-item requests by default (configurable
 from one to eight) and stores a small mapping plus a generation-named,
 validated `.hmls` mmap slab under codedb's local per-project data directory
 (0700/0600 on POSIX).
@@ -161,7 +161,7 @@ codedb-cli stop                       # stop daemon
 | `CODEDB_EMBEDDINGS_DIMENSIONS` | managed | Requested custom-provider dimensions (64-4096) |
 | `CODEDB_EMBEDDINGS_TOKEN` | unset | Optional legacy bearer token for protected/custom endpoints; the hosted lane enrolls automatically |
 | `CODEDB_EMBEDDINGS_TIMEOUT_MS` | `15000` | Per-request deadline in milliseconds (10-120000) |
-| `CODEDB_SEMANTIC_INDEX_CONCURRENCY` | `4` | Parallel 25-item index batches (clamped to 1-8) |
+| `CODEDB_SEMANTIC_INDEX_CONCURRENCY` | `8` | Parallel 100-item index batches (clamped to 1-8) |
 
 ## Performance
 

docs/releases/0.2.5860.md

@@ -0,0 +1,17 @@
+# CodeDB 0.2.5860
+
+Faster hosted semantic index builds.
+
+## Hosted semantic retrieval
+
+`codedb semantic-index` now sends 100-item batches with eight requests in flight by default. `CODEDB_SEMANTIC_INDEX_CONCURRENCY` still overrides the concurrency (1-8).
+
+Index builds wait out hosted rate limits with a doubling backoff capped at 30 seconds, instead of failing after a few quick retries. Interactive `codedb context` calls are unchanged and still fall back to local results immediately.
+
+Local retrieval, the watcher, and the installer are unchanged.
+
+## Verification
+
+```sh
+zig build test --summary all
+```

npm/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codedeebee",
-  "version": "0.2.5859",
+  "version": "0.2.5860",
   "description": "Zig code intelligence MCP server — npx launcher for the codedb native binary",
   "license": "MIT",
   "author": "justrach",

src/release_info.zig

@@ -1 +1 @@
-pub const semver = "0.2.5859";
+pub const semver = "0.2.5860";

src/semantic.zig

@@ -21,19 +21,22 @@ pub const default_timeout_ms: u32 = 15_000;
 pub const min_timeout_ms: u32 = 10;
 pub const max_timeout_ms: u32 = 120_000;
 pub const max_documents: usize = 24;
-// The free public codedb lane accepts at most 25 inputs per request. Indexing
-// may fill all 25 slots because, unlike query reranking, it does not prepend a
-// query vector to the batch.
-pub const max_index_documents: usize = 25;
+// The hosted lane accepts at most 100 inputs per request, the provider's own
+// batch ceiling. Larger batches cut per-item latency and round trips; query
+// reranking stays bounded separately by `max_documents`.
+pub const max_index_documents: usize = 100;
 pub const max_document_bytes: usize = 2 * 1024;
-// Keep the serialized anonymous request below Caddy's 64 KiB raw-body cap even
+// Keep the serialized request below the hosted edge's 256 KiB body cap even
 // when source punctuation/control bytes expand during JSON escaping.
 pub const max_batch_text_bytes: usize = 8 * 1024;
-pub const max_request_bytes: usize = 60 * 1024;
+pub const max_request_bytes: usize = 240 * 1024;
 pub const max_response_bytes: usize = 2 * 1024 * 1024;
-pub const max_index_batch_text_bytes: usize = 25 * 1024;
-pub const max_index_embedding_attempts: usize = 3;
+pub const max_index_batch_text_bytes: usize = 100 * 1024;
+// Hosted quota windows are per minute. Ten attempts with doubling backoff
+// capped at 30s wait about 92s in total, enough to outlast one window.
+pub const max_index_embedding_attempts: usize = 10;
 pub const index_embedding_retry_base_ms: u32 = 250;
+pub const index_embedding_retry_max_ms: u32 = 30_000;
 pub const default_rrf_k: f32 = 60;
 pub const default_semantic_weight: f32 = 0.05;
 // Accuracy-oriented Jina hybrid policy; see ADR 0008 for live-service
@@ -494,14 +497,32 @@ pub fn embedIndexRemoteTexts(
                 err == error.EmbeddingProviderUnavailable or
                 err == error.EmbeddingTimeout;
             if (!retryable or attempt + 1 == max_index_embedding_attempts) return err;
-            const delay_ms = index_embedding_retry_base_ms << @intCast(attempt);
+            const delay_ms = indexRetryDelayMs(attempt);
             io.sleep(.fromMilliseconds(delay_ms), .awake) catch {};
             continue;
         };
     }
     unreachable;
 }
 
+pub fn indexRetryDelayMs(attempt: usize) u32 {
+    const shift: u5 = @intCast(@min(attempt, 16));
+    const scaled = @as(u64, index_embedding_retry_base_ms) << shift;
+    return @intCast(@min(scaled, index_embedding_retry_max_ms));
+}
+
+test "index retry backoff doubles, caps, and outlasts a one-minute quota window" {
+    const testing = std.testing;
+    try testing.expectEqual(@as(u32, 250), indexRetryDelayMs(0));
+    try testing.expectEqual(@as(u32, 500), indexRetryDelayMs(1));
+    try testing.expectEqual(index_embedding_retry_max_ms, indexRetryDelayMs(8));
+    try testing.expectEqual(index_embedding_retry_max_ms, indexRetryDelayMs(40));
+    var total: u64 = 0;
+    for (0..max_index_embedding_attempts - 1) |attempt| total += indexRetryDelayMs(attempt);
+    try testing.expect(total >= 60_000);
+    try testing.expect(total <= 120_000);
+}
+
 pub fn embedQueryRemote(
     io: std.Io,
     allocator: std.mem.Allocator,

src/semantic_index.zig

@@ -26,12 +26,10 @@ pub const slab_file_suffix = ".hmls";
 pub const max_chunk_card_bytes: usize = 1024;
 pub const chunk_source_bytes: usize = 832;
 pub const max_chunks_per_file: usize = 256;
-// Hosted-lane hill climbs: 256 chunks (2026-08-26) gave c1=12.4s, c2=9.4s,
-// c4=6.1s, c8=16.7s. A 6,713-chunk React subset (2026-08-30) confirmed
-// c1=180.8s, c2=98.9s, c4=68.8s, while c8 was aborted after 338s in retry/
-// queue contention. Four remains the Pareto point; differently provisioned
-// endpoints can still use the explicit override.
-pub const default_parallel_batches: usize = 4;
+// Hosted 100-item batches (2026-09-27, 1,570 chunks): c4=10.7s, c8=5.9s
+// against 17.8s for the former 25-item c4 default. Repositories beyond the
+// hosted per-minute quota are paced by the index retry backoff instead.
+pub const default_parallel_batches: usize = 8;
 pub const max_parallel_batches: usize = 8;
 pub const default_search_results: usize = 24;
 pub const max_records: usize = 250_000;
@@ -1657,15 +1655,13 @@ test "semantic ANN replacement removes only the previously referenced slab" {
     try std.Io.Dir.cwd().access(io, new_path, .{});
 }
 
-test "semantic ANN wave uses the public 25-item batch boundary" {
-    // This is the hosted-lane Pareto point measured against the production
-    // TurboAPI/TEI queue. Raising the default to the supported maximum of 8
-    // regressed the 256-chunk end-to-end run from 6.1s to 16.7s. Keep the
-    // override for custom endpoints, but make an accidental default drift a
-    // release-gating test failure.
-    try std.testing.expectEqual(@as(usize, 4), default_parallel_batches);
-    try std.testing.expectEqual(@as(usize, 25), semantic.max_index_documents);
-    try std.testing.expectEqual(@as(usize, 100), default_parallel_batches * semantic.max_index_documents);
+test "semantic ANN wave uses the hosted 100-item batch boundary" {
+    // Hosted batches carry 100 items with eight in flight (800-item waves).
+    // Keep the override for custom endpoints, but make an accidental default
+    // drift a release-gating test failure.
+    try std.testing.expectEqual(@as(usize, 8), default_parallel_batches);
+    try std.testing.expectEqual(@as(usize, 100), semantic.max_index_documents);
+    try std.testing.expectEqual(@as(usize, 800), default_parallel_batches * semantic.max_index_documents);
     try std.testing.expectEqual(@as(usize, 1), embeddingBatchCount(semantic.max_index_documents));
     try std.testing.expectEqual(@as(usize, 2), embeddingBatchCount(semantic.max_index_documents + 1));
     try std.testing.expectEqual(