The new LLM extraction router shipped with little branch coverage, dropping src/nest below the 80% gate. Add unit tests for routeExtraction (flights/single/union/error paths, deterministic booking-wide fill), the native Ollama format client, the provider factory, the local-router service path with its type-aware text cap, the flat->schema.org mapper's remaining reservation types, and the background import-jobs runner. Also remove the now-unused validate.ts (only its FlatLike type was still referenced; moved to flat-schemas).
TripPlannerPage held a useEffect (the background-import → review bridge), which trips the page-pattern check (pages must stay wiring containers). Move the effect and its store/IndexedDB wiring into useTripPlanner where the rest of the import-review state already lives.
hasIntegrations gated the tab on memories/mcp/airtrail only, so a user with just the llm_parsing addon enabled saw no Integrations tab and could not reach the AI parsing config. Include llmEnabled in the gate.
Adds an "AI parsing" section under Settings -> Integrations where a user can choose the LLM provider, model, base URL, API key and multimodal option used for booking extraction. This per-user config applies when an admin has not configured an instance-wide model. Reuses the existing encrypted user settings: the API key is stored encrypted, never prefilled, and a blank field keeps the stored one. Adds settings.aiParsing.* across all 20 locales.
The source document was only kept in memory on the background task, so a page reload during the (now always-LLM ~25s) parse lost it and the booking saved without its file. Store the uploaded files in IndexedDB keyed by job id; the review loads them from there when the in-memory copy is gone, and a 1h TTL prunes abandoned imports.
During the per-item import review the booking isn't saved yet, so the Costs section showed an empty 'Create expense' even though a linked cost will be created on save. Show the parsed price (amount + category) as the pending linked expense so the user can verify it up front. Reuses existing i18n keys.
A long single-booking PDF (e.g. an 11-page rental voucher) spent ~200s on CPU prompt-eval at the 16k cap, though its data sits in the first ~2k. Cap non-flight docs at 6k (flights keep 16k for all legs). Also make the rental operator a required field so the car gets a real title.
The saving client gets no budget:created echo (X-Socket-Id) and the create response omits the linked budget item, so the booking's Costs section and the Costs tab stayed stale until a manual reload. Reload the budget items right after a create that carried a budget entry.
After dropping the vendor templates, the model skipped the (often unlabeled) Expedia-style hotel address — making address a required schema field forces it to emit the street-address line, restoring the booking's location/place. Also hint the rental company so a car booking gets a real title instead of the generic fallback.
Now that a capable instruct model (Qwen3-8B, thinking off) reads name/address/dates/legs reliably across formats, the per-vendor template short-circuit distorted more than it fixed: brittle on layout variations and overriding the better model output. Remove the template layer; the model extracts the structure and Schicht 2 backfills the confirmation/total and takes the currency from the document's own symbol (correcting model misreads like ¥→$). Per-type prompts now also ask for address and price/currency.
A/B against the prior default (qwen2.5:7b) on CPU showed Qwen3-8B is both faster and more accurate on tricky/multilingual booking docs (correct Airbnb year+price, correct DisneySea admission date), once thinking is disabled — which the router now does. Feature it as the recommended pull, keep qwen2.5:7b as the fallback.
Hybrid/reasoning models (Qwen3 and similar) default to emitting reasoning tokens, which collide with Ollama's format-grammar constraint — on CPU this produced null/unparseable output and blew the latency budget (qwen3:8b: null or 300s timeouts vs ~20s with thinking off). Send think:false on the /api/chat call; Ollama ignores it for non-thinking models (verified on qwen2.5:7b), so it's safe and unlocks the stronger Qwen3 family.
Keep the uploaded files on the background task and hand them to the review flow, so each reviewed booking pre-fills its Files with the document it was parsed from (uploaded with the booking on save). The two modals also adopt the shared resolveDayId helper.
Code-audit clean-ups: share one normCurrency between the router and the templates, lift the duplicated nearest-day resolver into formatters.resolveDayId, drop two needless as-unknown-as casts at the fillBookingWideFields call sites, restore routeExtraction's doc comment, and give the broker template readable names. Plus recognise ¥/JPY and fall back to a standalone symbol amount, so a Klook-style voucher whose price sits far from any label still yields a cost.
Imported bookings auto-create their linked budget items server-side, but the saving client suppresses its own budget:created echo, so the Costs list stayed stale until a manual reload. Reload the budget items when the review session ends.
A reviewed transport (e.g. a rental car) arrived with only its parsed pick-up/return dates and no day_id, so the modal kept just the time and saved a bare "HH:MM" with no date. Resolve start/end day from the parsed dates (exact match, else nearest trip day) so the booking lands on the right days.
reservation_endpoints.lat/lng are NOT NULL, so saving a reviewed transport whose pick-up/return couldn't be geocoded threw a 500 and lost the whole booking (dates, linked cost). Skip those rows; the dates still persist on reservation_time/reservation_end_time.
Persist the background-import tasks (id/trip/status only) and re-fetch each job's status on mount, so a parse still running when the page reloads keeps its widget instead of vanishing; expired jobs (404) are dropped and a restored 'done' task re-fetches its items.
Apply the deterministic confirmation-code and total fill to vendor-template results too (not just model output), and require the captured reference to contain a digit so a bare 'Confirmation'/'Reference' label no longer grabs the next prose word.
Pull the hotel/rental fields these vendors print in a stable text layout (name, address, stay/pickup dates, price, reference) deterministically, so the import stops depending on the local model for them. Handles German long/abbreviated months and English dates incl. 12-hour and comma forms.
Reviewing an imported booking saves it through the normal reservation
form, which dropped the parsed price (so no linked cost was created) and
only created the accommodation when both nights matched a trip day.
Carry the parsed price into a linked cost on save, and create the
accommodation from whichever day the check-in/out dates resolve to.
Parsing a booking can take a while on a CPU host, so don't hold the
upload modal open for it. The async import endpoint returns a job id
right away; the parse runs server-side (one at a time per user) and
pushes progress over the user's WebSocket, and a small widget in the
bottom corner tracks it while the user keeps navigating and editing.
A finished job opens the per-item review from the widget.
The single-shot prompt was unreliable on multi-leg flights and longer
documents, and slow on a CPU host. For the local provider, run a small
router instead:
- deterministic vendor templates first, with no model call at all
- exactly one grammar-enforced call per document via Ollama's native
`format` (flights as a flat array of legs, everything else as one flat
reservation, the type picked from keywords or a union schema)
- booking-wide fields (booking reference, total price, the overnight
arrival day) filled deterministically from the text afterwards, and
dates coerced to ISO so a natural-language date can't slip through
Recommend qwen2.5 in the AI-parsing settings instead of NuExtract.
Instead of writing parsed items straight to the trip, the import opens the
normal edit modal pre-filled for each one, so you can check and fix it before
saving — useful when a model guesses a wrong date or address. Hotels gained an
editable address field; on save an existing place is matched by name, otherwise
the reviewed address is geocoded and a new place is created.
The provider picker is the shared CustomSelect now and the form is split into
clear sections rather than a flat stack of inputs. NuExtract 2.0 2B is the
recommended default — fastest on a CPU-only host and MIT licensed; the 4B
carries a non-commercial licence, so it's no longer flagged as recommended.
NuExtract isn't an instruct model — fed a plain chat prompt it just echoes the
schema back. Detect a NuExtract model by id and talk to it the way the model
cards document: the JSON template inlined in a single user message, no system
prompt, no json_schema, temperature 0. Its flat result is mapped back to the
same KiReservation shape the rest of the pipeline already uses, so nothing
downstream changes; every other model keeps the generic prompt.
Money is taken as a verbatim string and parsed locally (German "1.580,22 €"
otherwise comes back as 1.49772), a rental car's pickup/return ride the from/to
fields so a stray form label doesn't become the location, and a lodging with no
name falls back to its address instead of being dropped.
A freshly imported hotel links to an accommodation that lives outside the
trip store, so loadTrip alone left the reservation edit modal with blank
place/date fields. Reload the accommodations list once the import finishes.
When a confirmation carries a total price, record it as a real expense
linked to the reservation (in the matching Costs category) instead of
leaving the amount in metadata only. Gated on the Costs addon.
Request and map root-level seat/class/platform and a total price/currency into reservation metadata (shown on the card; price reuses the existing label). Read both the root and reservationFor and tolerate common field-name aliases (priceAmount, priceCurrencyISO4217Code, fareClass, ...) since models name these inconsistently. Also capture event/attraction venue telephone + url onto the auto-created place, matching lodging/restaurant.
On a GPU-less host the model's prompt-eval time scales with input length and dominates total latency. Booking details sit at the top of a confirmation, so capping the extracted text at 4000 chars (was 8000) roughly halves extraction time (~50s warm for a capable local 7B model) with no loss of fields on real hotel/rental confirmations. Tunable if a long multi-segment itinerary needs more.
- rental car: request+map dropoffLocation, emit pickup->return from/to endpoints, set a location string (G1/G2/G3). - geocode endpoints (stations/stops/terminals/rental desks) on confirm via Nominatim; mapper now emits coordless named endpoints and confirm persists only the geocoded ones (G6). - assign every dated booking to the nearest trip day so it still shows when slightly out of range, and keep hotel accommodation from vanishing when a check date misses (G5/G10). - fix bus mislabelled as train + add bus_number metadata (G7/G8), flag malformed boats (G9), accept root start/end time for events (G11). - raise the local-LLM timeout to 300s for CPU-only Ollama.
client: the import call inherited the global 8s axios timeout and aborted long LLM extractions even though the server finished it; remove the timeout. server: raise the OpenAI-compatible LLM timeout 60s->180s (a cold Ollama model can take ~45s to first token). server: cap extracted text to 8000 chars before the LLM - multi-page T&C tails (30k+ chars) overflowed the context window, truncating the relevant head and making CPU inference crawl; booking details sit at the top.
Booking import only fell back to the LLM when each user flipped an 'always retry with AI' toggle, so by default files kitinerary returned nothing for just failed. Run the fallback automatically whenever the AI Parsing addon is on (fallback-on-empty); drop the now-redundant per-user toggle and its setting.
The map route ran first-activity to last-activity only, while the sidebar
already showed the hotel-to-first-stop and last-stop-to-hotel legs with
their drive times. Feed the day's accommodation bookends into the map
route too, reusing the same getDayBookendHotels lookup and the
"optimize from accommodation" gate, so the drawn line starts and ends at
the hotel, including single-activity and transfer days.
Adding an expense required at least one participant, so a cost you only
want to record — e.g. a booking paid on-site later — could not be saved
without splitting it. Drop the participant requirement: with nobody
selected the expense saves as a recorded total, counted in the trip
total and shown as Unfinished, and kept out of settlements until
who-paid is filled in. The shared schema and server already supported
this case.
Removing the only item of a user-created category deleted the whole
category. Turn that row back into the existing ... placeholder in
place instead, so the category keeps its position and colour; adding an
item reuses the placeholder slot. Deleting the placeholder (or the
category menu) still removes an empty category.
When the backend or identity provider was unreachable, a returning user with a
persisted session landed on the dashboard with an empty trip grid and no error.
That looks identical to a logged-in user who simply has no trips, so people
assumed their data had been lost.
Three client-side layers were quietly swallowing the failure: the auth check
only cleared state on a 401, so a 5xx or a network error left the stale session
in place and kept rendering the protected route; the offline-first trip repo
turned a failed fetch into the empty cache without throwing; and the dashboard
had neither an error nor an empty state, so a blank grid meant both "outage" and
"no trips".
The auth check now tells genuine offline (keep serving the cache silently, the
PWA happy path) apart from a server outage while online (keep the session but
flag it). The dashboard shows a reassuring "couldn't reach the server, your
trips are safe" banner with a retry, and a real zero-trip account finally gets a
proper empty state so the two cases never look alike. New strings added across
all locales.
On OIDC-only instances the bootstrap admin (first SSO user) rarely carries the configured admin claim, so a forced re-login — e.g. after a JWT-secret rotation — re-derived its role purely from claims and demoted it to user, locking the instance out with no recovery. The OIDC login role sync now skips a downgrade that would strip the last remaining admin, and the admin user-update endpoint guards the same case.
The day-plan route bar lost its Open in Google Maps action in the 3.1.0 redesign. A small button with the Google logo (monochrome, theme-aware) now sits next to the Route toggle and opens the day stops, in planned order, as a Google Maps directions link in a new tab.
The "How to Update" modal always rendered Docker commands and claimed the instance runs in Docker, even on bare-metal / LXC installs like Proxmox Community Scripts. It now branches on the is_docker flag the backend already returns: non-Docker installs get a generic "re-run your install method" note plus a link to the update guide. Docker stays the default when the flag is absent, so existing installs are unaffected.
Maurice