mirror of
https://github.com/tinyhumansai/openhuman.git
synced 2026-07-27 21:08:00 +00:00
fix(observability,embeddings): demote Ollama embed user-config rejections (Sentry TAURI-RUST-XS + MA/KM/GX) (#2612)
This commit is contained in:
oxoxDev
@@ -160,6 +160,17 @@ pub fn expected_error_kind(message: &str) -> Option<ExpectedErrorKind> {
|
||||
if is_loopback_unavailable(&lower) {
|
||||
return Some(ExpectedErrorKind::LoopbackUnavailable);
|
||||
}
|
||||
// Check `is_ollama_user_config_rejection` BEFORE the generic network /
|
||||
// backend-error matchers: the GX "daemon unreachable at localhost" shape
|
||||
// contains a loopback host but no `Connection refused (os error …)`
|
||||
// marker, and the XS / MA / KM 400/404 shapes are pure user-config —
|
||||
// wrong model name, model not pulled, daemon opted-in but not running.
|
||||
// Route them to the dedicated arm so they share the `ProviderUserState`
|
||||
// bucket with the composio / OAuth user-state errors instead of falling
|
||||
// through to capture. See `is_ollama_user_config_rejection`.
|
||||
if is_ollama_user_config_rejection(&lower) {
|
||||
return Some(ExpectedErrorKind::ProviderUserState);
|
||||
}
|
||||
if is_network_unreachable_message(&lower) {
|
||||
return Some(ExpectedErrorKind::NetworkUnreachable);
|
||||
}
|
||||
@@ -284,6 +295,77 @@ fn is_loopback_unavailable(lower: &str) -> bool {
|
||||
|| lower.contains("connection refused (os error 10061)")
|
||||
}
|
||||
|
||||
/// Detect Ollama embed call sites that surface a user-config rejection from
|
||||
/// the local Ollama daemon — pure user-state errors the UI already surfaces
|
||||
/// (toast / settings page warning) where Sentry has no remediation path.
|
||||
///
|
||||
/// Three canonical wire shapes are covered, all emitted by
|
||||
/// `openhuman::embeddings::ollama::OllamaEmbedding::embed` and the embed
|
||||
/// service fallback path:
|
||||
///
|
||||
/// - **TAURI-RUST-XS** (~376 events on self-hosted Sentry): user pointed the
|
||||
/// embedder at a chat / vision model id with a temperature suffix (e.g.
|
||||
/// `qwen3-vl:4b@0.7`) which Ollama parses as malformed. Wire shape:
|
||||
/// `ollama embed failed with status 400 Bad Request: {"error":"invalid model name"}`.
|
||||
/// - **OPENHUMAN-TAURI-MA / -KM** (deferred follow-up from PR #2216): user
|
||||
/// configured a model id that the local Ollama daemon hasn't pulled yet.
|
||||
/// Wire shape:
|
||||
/// `ollama embed failed with status 404 Not Found: {"error":"model \"<id>\" not found, try pulling it first"}`.
|
||||
/// - **OPENHUMAN-TAURI-GX**: user opted into Ollama embeddings but the
|
||||
/// daemon isn't running on `localhost:11434`, so the embed service falls
|
||||
/// back to cloud embeddings for the session. Wire shape:
|
||||
/// `ollama embeddings opted-in but daemon unreachable at http://localhost:11434; falling back to cloud embeddings for this session`.
|
||||
///
|
||||
/// All three are user-config: the user picked the wrong model id, forgot to
|
||||
/// pull it, or forgot to start the daemon. The remediation is "fix the
|
||||
/// model id in Settings" / "run `ollama pull <id>`" / "start ollama" —
|
||||
/// none of which Sentry can do for them.
|
||||
///
|
||||
/// The classifier is anchored on the `"ollama embed"` prefix
|
||||
/// (`"ollama embed failed"` for the 400/404 shapes, `"ollama embeddings opted-in"`
|
||||
/// for the daemon-unreachable fallback) so unrelated 400/404 errors elsewhere
|
||||
/// in the codebase that happen to contain `"invalid model name"` or
|
||||
/// `"not found"` substrings are not silenced.
|
||||
///
|
||||
/// Routes to [`ExpectedErrorKind::ProviderUserState`] — the same bucket that
|
||||
/// holds the composio / gmail / OAuth user-state errors. We deliberately do
|
||||
/// **not** introduce a dedicated Ollama enum variant: the demotion semantics
|
||||
/// (drop to `info` log, skip Sentry capture) are identical and adding a new
|
||||
/// variant for every provider would balloon the enum without changing
|
||||
/// behavior.
|
||||
fn is_ollama_user_config_rejection(lower: &str) -> bool {
|
||||
// XS — 400-status user-config (invalid model name, including the
|
||||
// temperature-suffix shape `qwen3-vl:4b@0.7` Ollama parses as malformed).
|
||||
if lower.contains("ollama embed failed") && lower.contains("invalid model name") {
|
||||
return true;
|
||||
}
|
||||
|
||||
// MA / KM — 404-status pull-required. The wire shape is JSON-escaped
|
||||
// (`\"<model-id>\" not found`); after lower-casing we still see the
|
||||
// backslash-quoted form. Anchor on `model \"` + `\" not found` so an
|
||||
// unrelated 404 that merely contains `"model"` and `"not found"` is not
|
||||
// swallowed. The `\\"` byte pair in Rust source matches the literal
|
||||
// `\"` sequence in the wire shape.
|
||||
if lower.contains("ollama embed failed")
|
||||
&& lower.contains("model \\\"")
|
||||
&& lower.contains("\\\" not found")
|
||||
{
|
||||
return true;
|
||||
}
|
||||
|
||||
// GX — daemon-unreachable opt-in state. The wire shape is emitted by
|
||||
// the embed service when the user has opted into Ollama in settings
|
||||
// but the daemon isn't responding, so the service falls back to cloud
|
||||
// embeddings for the session. Anchor on the full prefix to keep the
|
||||
// matcher from colliding with unrelated `"daemon unreachable"`
|
||||
// messages from other domains (e.g. backend connection-health logs).
|
||||
if lower.contains("ollama embeddings opted-in but daemon unreachable at") {
|
||||
return true;
|
||||
}
|
||||
|
||||
false
|
||||
}
|
||||
|
||||
/// Detect transport-level connection failures that fire before any HTTP status
|
||||
/// is observed — DNS resolution failures, TCP connect refused/reset, TLS
|
||||
/// handshake failures, or ISP/firewall blocks. The canonical shape is
|
||||
@@ -1280,6 +1362,71 @@ mod tests {
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn classifies_ollama_user_config_rejections() {
|
||||
// TAURI-RUST-XS (~376 events): user pointed embedder at a chat /
|
||||
// vision model id, sometimes with a temperature suffix like `@0.7`
|
||||
// that Ollama parses as malformed.
|
||||
for raw in [
|
||||
// Canonical XS wire shape from
|
||||
// `OllamaEmbedding::embed` non-2xx path on a 400 Bad Request.
|
||||
r#"ollama embed failed with status 400 Bad Request: {"error":"invalid model name"}"#,
|
||||
// Same shape with a temperature-suffix model id the user pasted
|
||||
// into Settings → Embeddings → Ollama.
|
||||
r#"ollama embed failed with status 400 Bad Request: {"error":"invalid model name: qwen3-vl:4b@0.7"}"#,
|
||||
// OPENHUMAN-TAURI-MA — model not pulled (404 Not Found).
|
||||
r#"ollama embed failed with status 404 Not Found: {"error":"model \"bge-m3\" not found, try pulling it first"}"#,
|
||||
// OPENHUMAN-TAURI-KM — same shape, different model id + `:latest` tag.
|
||||
r#"ollama embed failed with status 404 Not Found: {"error":"model \"nomic-embed-text:latest\" not found, try pulling it first"}"#,
|
||||
// OPENHUMAN-TAURI-GX — daemon-unreachable opt-in state.
|
||||
"ollama embeddings opted-in but daemon unreachable at http://localhost:11434; falling back to cloud embeddings for this session",
|
||||
] {
|
||||
assert_eq!(
|
||||
expected_error_kind(raw),
|
||||
Some(ExpectedErrorKind::ProviderUserState),
|
||||
"should classify Ollama user-config rejection: {raw}"
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn does_not_classify_unrelated_ollama_errors_as_user_config() {
|
||||
// Unrelated 500 — server-side ollama bug must still reach Sentry.
|
||||
assert_eq!(
|
||||
expected_error_kind("ollama embed failed with status 500"),
|
||||
None
|
||||
);
|
||||
// Parse-failure on the response — real bug in either the server
|
||||
// or our deserializer, must still reach Sentry.
|
||||
assert_eq!(
|
||||
expected_error_kind(
|
||||
"ollama embed response parse failed: invalid type: expected sequence"
|
||||
),
|
||||
None
|
||||
);
|
||||
// Dimension mismatch — real bug (model dims don't match what we
|
||||
// recorded), must still reach Sentry.
|
||||
assert_eq!(
|
||||
expected_error_kind(
|
||||
"ollama embed dimension mismatch at index 0: expected 768, got 1024"
|
||||
),
|
||||
None
|
||||
);
|
||||
// Unrelated `invalid model name` outside Ollama embed call —
|
||||
// anchor on the `ollama embed` prefix keeps this from being silenced.
|
||||
assert_eq!(
|
||||
expected_error_kind("provider config validation failed: invalid model name"),
|
||||
None
|
||||
);
|
||||
// Unrelated `model "…" not found` text without the `ollama embed`
|
||||
// prefix — anchor keeps this from being silenced even when the
|
||||
// exact MA/KM wire-shape substring appears in another context.
|
||||
assert_eq!(
|
||||
expected_error_kind(r#"provider listing failed: model \"foo\" not found in registry"#),
|
||||
None
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn classifies_local_ai_capability_unavailable_errors() {
|
||||
// OPENHUMAN-TAURI-3B: surfaced by `local_ai_download_asset` when a
|
||||
|
||||
@@ -327,32 +327,60 @@ async fn embed_connection_refused() {
|
||||
);
|
||||
}
|
||||
|
||||
// OPENHUMAN-TAURI-{GP,MA,KM,GX} wire shapes — currently routed through
|
||||
// `report_error_or_expected` (Sentry classifier ladder) by this PR. The ladder
|
||||
// matches GP (LocalAiCapabilityUnavailable) today; MA/KM/GX still fall through
|
||||
// to capture because `observability::expected_error_kind` has no matcher arm
|
||||
// for "ollama model not found" / "ollama daemon unreachable". Those matcher
|
||||
// arms are blocked behind PR #2063 + #2188 merging (both touch
|
||||
// `src/core/observability.rs`) and will land in the follow-up classifier
|
||||
// batch. Tests below lock the CURRENT state so the follow-up flips them.
|
||||
// OPENHUMAN-TAURI-{GP,MA,KM,GX} + TAURI-RUST-XS wire shapes — routed through
|
||||
// `report_error_or_expected` (Sentry classifier ladder) and demoted to
|
||||
// `ProviderUserState` / `LocalAiCapabilityUnavailable` by the
|
||||
// `is_ollama_user_config_rejection` matcher in `src/core/observability.rs`.
|
||||
// GP catches the RAM-tier disable shape; XS/MA/KM/GX cover the four Ollama
|
||||
// user-config rejection shapes (400 invalid-model-name, 404 model-not-found
|
||||
// ×2, daemon-unreachable opt-in state).
|
||||
|
||||
#[test]
|
||||
fn ma_wire_shape_current_state_unclassified() {
|
||||
let msg = r#"ollama embed failed with status 404 Not Found: {"error":"model \"bge-m3\" not found, try pulling it first"}"#;
|
||||
fn xs_wire_shape_classifies_as_provider_user_state() {
|
||||
// TAURI-RUST-XS (~376 events on self-hosted Sentry): user pointed
|
||||
// embedder at a chat / vision model with a temperature suffix
|
||||
// (`qwen3-vl:4b@0.7`) Ollama parses as malformed.
|
||||
let msg = r#"ollama embed failed with status 400 Bad Request: {"error":"invalid model name"}"#;
|
||||
assert_eq!(
|
||||
crate::core::observability::expected_error_kind(msg),
|
||||
None,
|
||||
"MA — matcher arm pending follow-up classifier batch (post #2063 + #2188 merge)"
|
||||
Some(crate::core::observability::ExpectedErrorKind::ProviderUserState),
|
||||
"XS — invalid model name 400 is pure user-config, must demote"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn km_wire_shape_current_state_unclassified() {
|
||||
fn xs_temperature_suffix_model_classifies_as_provider_user_state() {
|
||||
// Same XS shape with the temperature-suffix model id embedded in the
|
||||
// body. Real wire shape from the field report.
|
||||
let msg = r#"ollama embed failed with status 400 Bad Request: {"error":"invalid model name: qwen3-vl:4b@0.7"}"#;
|
||||
assert_eq!(
|
||||
crate::core::observability::expected_error_kind(msg),
|
||||
Some(crate::core::observability::ExpectedErrorKind::ProviderUserState),
|
||||
"XS — temperature-suffix model id must also demote"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn ma_wire_shape_classifies_as_provider_user_state() {
|
||||
// OPENHUMAN-TAURI-MA: user configured an Ollama model id that the
|
||||
// local daemon hasn't pulled yet.
|
||||
let msg = r#"ollama embed failed with status 404 Not Found: {"error":"model \"bge-m3\" not found, try pulling it first"}"#;
|
||||
assert_eq!(
|
||||
crate::core::observability::expected_error_kind(msg),
|
||||
Some(crate::core::observability::ExpectedErrorKind::ProviderUserState),
|
||||
"MA — model-not-found 404 is pure user-state (pull required), must demote"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn km_wire_shape_classifies_as_provider_user_state() {
|
||||
// OPENHUMAN-TAURI-KM: same wire shape as MA with a different model
|
||||
// id + `:latest` tag.
|
||||
let msg = r#"ollama embed failed with status 404 Not Found: {"error":"model \"nomic-embed-text:latest\" not found, try pulling it first"}"#;
|
||||
assert_eq!(
|
||||
crate::core::observability::expected_error_kind(msg),
|
||||
None,
|
||||
"KM — matcher arm pending follow-up classifier batch"
|
||||
Some(crate::core::observability::ExpectedErrorKind::ProviderUserState),
|
||||
"KM — pull-required 404 must demote"
|
||||
);
|
||||
}
|
||||
|
||||
@@ -363,17 +391,30 @@ fn gp_wire_shape_classifies() {
|
||||
assert_eq!(
|
||||
crate::core::observability::expected_error_kind(msg),
|
||||
Some(crate::core::observability::ExpectedErrorKind::LocalAiCapabilityUnavailable),
|
||||
"GP — LocalAiCapabilityUnavailable matcher must catch this; closed by this PR"
|
||||
"GP — LocalAiCapabilityUnavailable matcher must catch this"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn gx_wire_shape_current_state_unclassified() {
|
||||
fn gx_wire_shape_classifies_as_provider_user_state() {
|
||||
// OPENHUMAN-TAURI-GX: user opted into Ollama embeddings in Settings
|
||||
// but the daemon isn't running on localhost:11434.
|
||||
let msg = "ollama embeddings opted-in but daemon unreachable at http://localhost:11434; falling back to cloud embeddings for this session";
|
||||
assert_eq!(
|
||||
crate::core::observability::expected_error_kind(msg),
|
||||
Some(crate::core::observability::ExpectedErrorKind::ProviderUserState),
|
||||
"GX — daemon-unreachable opt-in state is pure user-config (start daemon)"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn ollama_500_wire_shape_stays_unexpected() {
|
||||
// Server-side ollama bug — must still reach Sentry.
|
||||
let msg = "ollama embed failed with status 500 Internal Server Error: model crashed";
|
||||
assert_eq!(
|
||||
crate::core::observability::expected_error_kind(msg),
|
||||
None,
|
||||
"GX — matcher arm pending follow-up classifier batch"
|
||||
"real ollama server errors must still reach Sentry"
|
||||
);
|
||||
}
|
||||
|
||||
@@ -387,6 +428,19 @@ fn ollama_parse_error_wire_shape_stays_unexpected() {
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn ollama_dimension_mismatch_stays_unexpected() {
|
||||
// Dimension mismatch — real bug (model dims don't match what we
|
||||
// recorded for the provider in Settings), must still reach Sentry
|
||||
// so we can investigate the desync.
|
||||
let msg = "ollama embed dimension mismatch at index 0: expected 768, got 1024";
|
||||
assert_eq!(
|
||||
crate::core::observability::expected_error_kind(msg),
|
||||
None,
|
||||
"dimension-mismatch shape must still reach Sentry"
|
||||
);
|
||||
}
|
||||
|
||||
// ── embed_one (trait default) ───────────────────────────
|
||||
|
||||
#[tokio::test]
|
||||
|
||||
Reference in New Issue
Block a user