Skip to content

semantic: use negotiated float16 transport #730

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
justrach merged 1 commit into from
Aug 28, 2026
+168 −78
Merged

semantic: use negotiated float16 transport #730

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
181 changes: 103 additions & 78 deletions src/semantic.zig
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -241,6 +241,45 @@ const EmbeddingRequest = struct {
encoding_format: []const u8 = "float",
};

const compact_encoding = "base64-f16";

fn embeddingEncoding(config: Config) []const u8 {
// Keep custom OpenAI-compatible endpoints on the ubiquitous JSON-float
// contract. Only the CodeDB-operated endpoint negotiates the compact wire
// representation whose exact binary layout we control.
return if (config.isCodedbHosted()) compact_encoding else "float";
}

fn fetchEmbeddingInputs(
io: std.Io,
allocator: std.mem.Allocator,
config: Config,
inputs: []const []const u8,
) ![]u8 {
const encoding_format = embeddingEncoding(config);
const body = try std.json.Stringify.valueAlloc(allocator, EmbeddingRequest{
.model = config.model,
.input = inputs,
.dimensions = config.dimensions,
.encoding_format = encoding_format,
}, .{});
defer allocator.free(body);
return fetchEmbeddingResponse(io, allocator, config, body) catch |err| {
// A pre-negotiation CodeDB edge rejects the new request field with a
// 4xx. Retry once with the legacy contract and a fresh signed proof.
// Custom providers never enter this path.
if (!config.isCodedbHosted() or err != error.EmbeddingProviderRejected) return err;
const legacy_body = try std.json.Stringify.valueAlloc(allocator, EmbeddingRequest{
.model = config.model,
.input = inputs,
.dimensions = config.dimensions,
.encoding_format = "float",
}, .{});
defer allocator.free(legacy_body);
return fetchEmbeddingResponse(io, allocator, config, legacy_body);
};
}

fn fetchHttpResponseOnce(
io: std.Io,
allocator: std.mem.Allocator,
Expand Down Expand Up @@ -383,13 +422,7 @@ pub fn embedRemoteTexts(
}
if (text_bytes > max_index_batch_text_bytes) return error.EmbeddingRequestTooLarge;

const body = try std.json.Stringify.valueAlloc(allocator, EmbeddingRequest{
.model = config.model,
.input = inputs,
.dimensions = config.dimensions,
}, .{});
defer allocator.free(body);
const response = try fetchEmbeddingResponse(io, allocator, config, body);
const response = try fetchEmbeddingInputs(io, allocator, config, inputs);
defer allocator.free(response);
const vectors = try vectorsFromProviderResponse(allocator, response, inputs.len, config.dimensions, config.model);
return .{
Expand Down Expand Up @@ -534,15 +567,7 @@ pub fn scoreRemote(
inputs[i + 1] = text_storage.items[offset.start .. offset.start + offset.len];
}

const body = try std.json.Stringify.valueAlloc(allocator, EmbeddingRequest{
.model = config.model,
.input = inputs,
.dimensions = config.dimensions,
}, .{});
defer allocator.free(body);
if (body.len > max_request_bytes) return error.EmbeddingRequestTooLarge;

const response = try fetchEmbeddingResponse(io, allocator, config, body);
const response = try fetchEmbeddingInputs(io, allocator, config, inputs);
defer allocator.free(response);

const scores = try scoresFromProviderResponse(
Expand Down Expand Up @@ -587,6 +612,56 @@ fn validateProviderModel(root: std.json.Value, expected_model: ?[]const u8) !voi
if (model != .string or !std.mem.eql(u8, model.string, expected)) return error.EmbeddingModelMismatch;
}

const ResponseEncoding = enum { float, base64_f16 };

fn responseEncoding(root: std.json.Value) !?ResponseEncoding {
if (root != .object) return error.InvalidEmbeddingResponse;
const value = root.object.get("encoding_format") orelse return null;
if (value != .string) return error.InvalidEmbeddingResponse;
if (std.mem.eql(u8, value.string, "float")) return .float;
if (std.mem.eql(u8, value.string, compact_encoding)) return .base64_f16;
return error.InvalidEmbeddingResponse;
}

fn decodeEmbeddingValue(
row: []f32,
embedding_value: std.json.Value,
declared_encoding: ?ResponseEncoding,
) !void {
switch (embedding_value) {
.array => |array| {
if (declared_encoding == .base64_f16) return error.InvalidEmbeddingResponse;
if (array.items.len != row.len) return error.InvalidEmbeddingDimensions;
for (array.items, 0..) |value, i| {
const number = try numberAsF64(value);
if (!std.math.isFinite(number)) return error.InvalidEmbeddingNumber;
const narrowed: f32 = @floatCast(number);
if (!std.math.isFinite(narrowed)) return error.InvalidEmbeddingNumber;
row[i] = narrowed;
}
},
.string => |encoded| {
if (declared_encoding == .float) return error.InvalidEmbeddingResponse;
const byte_len = std.math.mul(usize, row.len, 2) catch return error.InvalidEmbeddingDimensions;
if (byte_len > 4096 * 2) return error.InvalidEmbeddingDimensions;
const decoded_len = std.base64.standard.Decoder.calcSizeForSlice(encoded) catch
return error.InvalidEmbeddingResponse;
if (decoded_len != byte_len) return error.InvalidEmbeddingDimensions;
var decoded_storage: [4096 * 2]u8 = undefined;
const decoded = decoded_storage[0..byte_len];
std.base64.standard.Decoder.decode(decoded, encoded) catch return error.InvalidEmbeddingResponse;
for (row, 0..) |*value, i| {
const bits = std.mem.readInt(u16, decoded[i * 2 ..][0..2], .little);
const half: f16 = @bitCast(bits);
const narrowed: f32 = @floatCast(half);
if (!std.math.isFinite(narrowed)) return error.InvalidEmbeddingNumber;
value.* = narrowed;
}
},
else => return error.InvalidEmbeddingResponse,
}
}

fn vectorsFromProviderResponse(
allocator: std.mem.Allocator,
json_text: []const u8,
Expand All @@ -598,6 +673,7 @@ fn vectorsFromProviderResponse(
defer parsed.deinit();
try validateProviderModel(parsed.value, expected_model);
if (parsed.value != .object) return error.InvalidEmbeddingResponse;
const declared_encoding = try responseEncoding(parsed.value);
const data_value = parsed.value.object.get("data") orelse return error.InvalidEmbeddingResponse;
if (data_value != .array or data_value.array.items.len != count) return error.InvalidEmbeddingResponse;

Expand All @@ -615,52 +691,12 @@ fn vectorsFromProviderResponse(
if (index >= count or seen[index]) return error.InvalidEmbeddingResponse;
seen[index] = true;
const embedding_value = item_value.object.get("embedding") orelse return error.InvalidEmbeddingResponse;
if (embedding_value != .array or embedding_value.array.items.len != dimensions) return error.InvalidEmbeddingDimensions;
const row = vectors[index * dimensions ..][0..dimensions];
for (embedding_value.array.items, 0..) |value, i| {
const number = try numberAsF64(value);
if (!std.math.isFinite(number)) return error.InvalidEmbeddingNumber;
const narrowed: f32 = @floatCast(number);
// JSON numbers are parsed as f64. Values such as 1e100 are finite
// there but overflow to infinity in the f32 index representation.
if (!std.math.isFinite(narrowed)) return error.InvalidEmbeddingNumber;
row[i] = narrowed;
}
try decodeEmbeddingValue(row, embedding_value, declared_encoding);
}
return vectors;
}

fn cosine(a: []const std.json.Value, b: []const std.json.Value) !f32 {
if (a.len == 0 or a.len != b.len) return error.InvalidEmbeddingDimensions;
var dot: f64 = 0;
var norm_a: f64 = 0;
var norm_b: f64 = 0;
for (a, b) |av, bv| {
const x = try numberAsF64(av);
const y = try numberAsF64(bv);
if (!std.math.isFinite(x) or !std.math.isFinite(y)) return error.InvalidEmbeddingNumber;
dot += x * y;
norm_a += x * x;
norm_b += y * y;
// Large-but-finite JSON scalars can overflow these f64 accumulators.
// Reject immediately rather than allowing NaN to be clamped into a
// plausible (and potentially perfect) retrieval score.
if (!std.math.isFinite(dot) or !std.math.isFinite(norm_a) or !std.math.isFinite(norm_b)) {
return error.InvalidEmbeddingNumber;
}
}
if (norm_a == 0 or norm_b == 0) return error.InvalidEmbeddingVector;
const norm_product = norm_a * norm_b;
if (!std.math.isFinite(norm_product) or norm_product <= 0) return error.InvalidEmbeddingNumber;
const denominator = @sqrt(norm_product);
if (!std.math.isFinite(denominator) or denominator == 0) return error.InvalidEmbeddingNumber;
const raw = dot / denominator;
if (!std.math.isFinite(raw)) return error.InvalidEmbeddingNumber;
const narrowed: f32 = @floatCast(@max(-1.0, @min(1.0, raw)));
if (!std.math.isFinite(narrowed)) return error.InvalidEmbeddingNumber;
return narrowed;
}

/// Parse an OpenAI-compatible embeddings response and return cosine scores for
/// document indices 1..N against query index 0. Items may arrive out of order.
pub fn scoresFromResponse(
Expand All @@ -679,31 +715,20 @@ fn scoresFromProviderResponse(
dimensions: u16,
expected_model: ?[]const u8,
) ![]f32 {
var parsed = try std.json.parseFromSlice(std.json.Value, allocator, json_text, .{});
defer parsed.deinit();
try validateProviderModel(parsed.value, expected_model);
if (parsed.value != .object) return error.InvalidEmbeddingResponse;
const data_value = parsed.value.object.get("data") orelse return error.InvalidEmbeddingResponse;
if (data_value != .array or data_value.array.items.len != document_count + 1) return error.InvalidEmbeddingResponse;

const vectors = try allocator.alloc(?[]const std.json.Value, document_count + 1);
const vectors = try vectorsFromProviderResponse(
allocator,
json_text,
document_count + 1,
dimensions,
expected_model,
);
defer allocator.free(vectors);
@memset(vectors, null);
for (data_value.array.items) |item_value| {
if (item_value != .object) return error.InvalidEmbeddingResponse;
const index_value = item_value.object.get("index") orelse return error.InvalidEmbeddingResponse;
if (index_value != .integer or index_value.integer < 0) return error.InvalidEmbeddingResponse;
const index: usize = @intCast(index_value.integer);
if (index >= vectors.len or vectors[index] != null) return error.InvalidEmbeddingResponse;
const embedding_value = item_value.object.get("embedding") orelse return error.InvalidEmbeddingResponse;
if (embedding_value != .array or embedding_value.array.items.len != dimensions) return error.InvalidEmbeddingDimensions;
vectors[index] = embedding_value.array.items;
}
const query = vectors[0] orelse return error.InvalidEmbeddingResponse;
const query = vectors[0..dimensions];
const scores = try allocator.alloc(f32, document_count);
errdefer allocator.free(scores);
for (scores, 0..) |*score, i| {
score.* = try cosine(query, vectors[i + 1] orelse return error.InvalidEmbeddingResponse);
const start = (i + 1) * dimensions;
score.* = try cosineF32(query, vectors[start..][0..dimensions]);
}
return scores;
}
Expand Down
65 changes: 65 additions & 0 deletions src/test_semantic.zig
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -202,6 +202,70 @@ test "semantic: parses raw vectors in response index order" {
try testing.expectEqualSlices(f32, &.{ 1, 0, 0, 1 }, vectors);
}

test "semantic: decodes negotiated little-endian base64 float16 vectors" {
const vectors = try semantic.vectorsFromResponse(
testing.allocator,
\\{"encoding_format":"base64-f16","data":[
\\ {"index":1,"embedding":"ADgAwA=="},
\\ {"index":0,"embedding":"ADwAAA=="}
\\]}
,
2,
2,
);
defer testing.allocator.free(vectors);
try testing.expectEqualSlices(f32, &.{ 1, 0, 0.5, -2 }, vectors);
}

test "semantic: float16 and JSON transports preserve retrieval ordering" {
const legacy =
\\{"data":[
\\ {"index":0,"embedding":[1,0]},
\\ {"index":1,"embedding":[0.75,0.25]},
\\ {"index":2,"embedding":[0,1]}
\\]}
;
const compact =
\\{"encoding_format":"base64-f16","data":[
\\ {"index":2,"embedding":"AAAAPA=="},
\\ {"index":0,"embedding":"ADwAAA=="},
\\ {"index":1,"embedding":"ADoANA=="}
\\]}
;
const legacy_scores = try semantic.scoresFromResponse(testing.allocator, legacy, 2, 2);
defer testing.allocator.free(legacy_scores);
const compact_scores = try semantic.scoresFromResponse(testing.allocator, compact, 2, 2);
defer testing.allocator.free(compact_scores);
try testing.expect(legacy_scores[0] > legacy_scores[1]);
try testing.expect(compact_scores[0] > compact_scores[1]);
try testing.expectApproxEqAbs(legacy_scores[0], compact_scores[0], 0.001);
try testing.expectApproxEqAbs(legacy_scores[1], compact_scores[1], 0.001);
}

test "semantic: compact decoder rejects malformed length and non-finite values" {
try testing.expectError(error.InvalidEmbeddingDimensions, semantic.vectorsFromResponse(
testing.allocator,
\\{"encoding_format":"base64-f16","data":[{"index":0,"embedding":"AAA="}]}
,
1,
2,
));
try testing.expectError(error.InvalidEmbeddingNumber, semantic.vectorsFromResponse(
testing.allocator,
\\{"encoding_format":"base64-f16","data":[{"index":0,"embedding":"AH4="}]}
,
1,
1,
));
try testing.expectError(error.InvalidEmbeddingResponse, semantic.vectorsFromResponse(
testing.allocator,
\\{"encoding_format":"float","data":[{"index":0,"embedding":"ADw="}]}
,
1,
1,
));
}

test "semantic: advisory fusion keeps lexical channel dominant" {
const best_lexical_worst_semantic = semantic.advisoryRrfScore(0, 23);
const second_lexical_best_semantic = semantic.advisoryRrfScore(1, 0);
Expand Down Expand Up @@ -366,6 +430,7 @@ test "semantic: endpoint override batches 25 inputs with auth and validates prov
defer parsed.deinit();
try testing.expectEqualStrings("mock-model", parsed.value.object.get("model").?.string);
try testing.expectEqual(@as(usize, 25), parsed.value.object.get("input").?.array.items.len);
try testing.expectEqualStrings("float", parsed.value.object.get("encoding_format").?.string);
}

test "semantic: provider model mismatch fails closed" {
Expand Down