Search engine runs in production mode with hard memory and CPU limits

The search engine now has a fixed memory ceiling, a CPU ceiling and a bounded indexing budget, so a heavy index update can never take memory away from the rest of the platform. Search results and the search APIs are unchanged; search briefly fell back to basic matching for about 30 seconds while the change was applied on 18 October.

Runaway jobs can no longer exhaust the server's memory

On 18 October the platform was unavailable twice, for about 10 minutes each time (around 10:53 and 11:54 Amman time), when a runaway non-production test process used up all of the server's memory. Non-production workloads now run under a hard memory cap, and a hardware watchdog restarts the server automatically if it ever stops responding.

Atlas answers only map and place questions

POST /ai/atlas now declines requests that have nothing to do with places, routes, traffic or the map, such as poems, homework, general trivia or unrelated code, with one friendly sentence in the user's language and action "answer". Questions about the place or area in view, including its history and what it is known for, are still answered.

Documentation assistant sticks to building on Navvo

The assistant on these docs now helps only with Navvo's APIs, SDKs, keys and docs, including code that calls Navvo, and points you to a relevant docs topic for anything else.

Traffic history is filed under the hour each drive happened

Contributed trips used to be recorded under the hour they were processed, so a trip uploaded late landed in the wrong hour of the learned history. Each road a trip drove is now recorded under the hour its GPS points were captured, and a trip that crosses an hour boundary is split by road. Profiles, insights and forecasts learn from the right hours from now on.

Traffic calendar follows each country's own weekend, Ramadan and holidays

The holiday, weekend and Ramadan flags behind traffic profiles now cover 15 countries, each in its own local time, with its own weekend for the period (Saturday–Sunday in the UAE and Lebanon, for example), Ramadan from the Hijri date, and its public holidays and Eids. Roads are matched to the country they are in instead of all being read as Jordan, and older dates are classified correctly instead of being treated as ordinary workdays.

Traffic history backfill runs once per restart

The background job that rebuilds traffic history after a restart was queued once per API worker. It is now queued once.

Expected traffic for any place and time in the Middle East

New `GET /traffic/analysed` returns the traffic expected at a point and moment, from Navvo's model of past traffic in the region, adjusted for Eid, public holidays and school holidays, the hour after iftar in Ramadan, the month, and heavy rain, storms or dust (from the forecast for the next 72 hours, or a what-if `weather`). `/traffic/analysed/week` adds seven days by hour with each day's best and worst two-hour windows, `/traffic/analysed/outlook` a short daily outlook in English or Arabic, and `/traffic/analysed/model` the model fits behind them. Answers are labelled `navvo-analysed-traffic`, are never live, and return `basis: "none"` instead of an error where the model has no data yet.

Wrong Arabic names removed from 1,076 places

An August backfill had given places the bare Arabic name of a brand mentioned in their English name, such as an ATM named as the whole bank or a shop named after its mall, so many unrelated places shared one Arabic name. Those 1,076 names are removed, and these places show their English name until a real Arabic one is added. The 140 places whose name is the brand keep it, and 6 got a better reading.

A chain's far branch no longer outranks nearby ones

With an origin, an exact name match more than 25 km away now competes on distance with nearer places whose names contain the query. So `carrefour` from Amman lists Amman's branches before one in another country. A major place, with thousands of reviews or very high popularity, keeps its lead however far it is, such as Petra or Al-Aqsa Mosque.

Web, Node, Python, Go and Java SDKs 0.3.0: price filter and result details

All five SDKs take the search `price` filter (levels 1–4) and return the new result details on search and similar-places results: today's hours, price range and level, a short address, area, review count, category name and highlights. Similar places also take a `locale`, and the Web SDK falls back to the client's `language`, so its `similar()` calls now send that language. Existing calls keep working, and downloads and checksums are at docs.navvo.io/sdks.

The script-tag Web SDK bundle now serves the latest version

The bundle the quickstart loads from `docs.navvo.io/navvo-maps/` was still version 0.1.0, so pages using it had none of the search features added since. It now serves 0.3.0, which keeps every 0.1.0 class and method. If such a page sends its requests through your own proxy, allow the `x-navvo-source` request header in the proxy's CORS policy.

Map search: only the search's pins stay on the map

While a search is shown, the map hides its other place pins, so only the results remain: the places and addresses it found, or the result you opened, and no pins at all when it finds nothing. They come back when the search ends, when you plan a route, or when you open a place another way. Your location, a dropped or parked pin, route warnings and the layers you switched on stay visible.

Map: an open place keeps its pin

A place opened from the map, a nearby list or a link could lose its pin once the map zoomed in to it, when another place stood close by. It now always keeps its pin, in the 3D view too, and the places crowding it step aside.

The first meaning-based search after a restart no longer comes back empty

Right after the search engine restarted, the first meaning-based query could miss its time limit while the engine loaded its data, and the request fell back to keyword-only results — for a descriptive query such as `place to watch football`, an empty list. The platform now runs that first query itself at start-up and every few minutes. A worker that cannot reach the engine also checks again after 30 seconds instead of treating meaning-based matching as unavailable for five minutes.

Map search: category chips and panel headings in Arabic

In the Arabic interface, the category chips (Restaurants, Cafés, Hotels and so on) and the panel's Explore, Recent and Popular nearby headings were still in English. They now show in Arabic and run right to left.

Search analytics missed two days of data (2–4 October)

Search analytics was not recorded from 2 October 02:55 to 4 October 11:02 Amman time. Collection pauses by design when the database disk has less than 3 GB free, and it stayed below that until space was freed on 4 October. Searches were answered normally throughout; searches and click events from that window are missing from analytics and cannot be recovered. Collection has resumed.

The search monitor now tests location search

Every five minutes the monitor takes a sample of places and checks that each one can be found by a radius search around its own position, and it notifies administrators after two failed runs. On 1 October every existing check stayed green while nearby search returned nothing for eleven hours; this check would have raised the alarm within ten minutes.

The standby search index is healthy again

The standby copy of the search index, left damaged by the 1 October repair, has been rebuilt and verified, so an administrator can switch to it at once if the live index is ever damaged.

`openNow` and opening hours work again

Open/closed state was not being computed for any place: the stored opening hours were in a format the platform did not read. `openNow=true` on `GET /search` and `GET /places` returned nothing, `GET /places/{slug}/hours` reported no hours, and place details carried no `openState`. Hours are read again for all 43,864 places that have them; a day with no entry stays unknown rather than being reported as closed.

Nearby and distance-sorted search results are complete again

On 1 October, from about 03:00 to 14:30 Amman time, searches that depend on distance — near-me queries, `radiusKm`, `sort=distance` and category searches around a location — could miss places close to the caller, because an overnight maintenance step damaged the location part of the search index. Searches by name were not affected. The index was rebuilt and the maintenance step now preserves location data.

Meaning-based matching is on

Searches that keyword matching cannot answer, such as `fix my phone screen`, `place to watch football` or `take my dog to the vet`, now also find places that fit the description, across Arabic and English, entirely on Navvo's own infrastructure. It only adds results when keyword matching finds few strong ones, never runs while someone is typing (`mode=typeahead`), and responses where it contributed report `index.mode: hybrid`. API-key requests opt in with `semantic=on`.

Meaning-based matching ignores random letters and reads Arabic as typed

Meaning-based results need some evidence in the query (a word that matches the catalogue, a category or a filter), so random letters and off-topic questions return nothing, and Arabic queries are matched in the spelling the user typed. The hourly meaning-based budget for requests without an API key now counts only searches in which it actually runs.

An index swap now checks nearby search first

Before a rebuilt search index replaces the live one, the swap now samples places from it and confirms that each can be found by location, not only that every place is present. The copy left over from the 1 October repair held every place but answered no nearby search; switching to it is now refused with a request to rebuild first.

Platform files consolidated into the platform's own account

Everything the platform writes — logs, build and QA artifacts, scratch work — now lives inside its own account and is owned by it, instead of being scattered across shared system locations. 1.1 GB of old load-test logs moved off a nearly full system disk, 5,165 files that had been created with administrator ownership were handed back to the account, and an hourly check now keeps it that way. The nightly database backup and the weekly data refresh keep their temporary files inside the account too. No endpoint or response changed.

Atlas answers on a faster, larger model

The model behind the map assistant is now a larger hosted model that typically answers in about one second, including when it looks up places or plans a route, and writes natural Arabic. The request and response contract is unchanged — the same actions, the same fields, and `model` is still always `navvo-ai`. If the hosted model is unavailable, Atlas falls back to a local model and then to its deterministic parser, as before.

Atlas no longer presents usual traffic as live

Most of Navvo's traffic answers are the usual pattern for the current hour, learned from past trips, rather than a live reading. Atlas used to receive that answer without its basis, so it could tell a user that a road was free flowing "right now". It now receives the basis with the data and phrases the answer accordingly — "usually at this time of day" — and makes clear that the answer covers an area, not a single road. Atlas also states only facts that came from Navvo's data in the same conversation, and says so when that data does not cover something, such as opening hours.

"Near me" and traffic questions use the user's own area

When a user asked for places near them, Atlas searched without the user's position, so results were not ranked by proximity; and "how is traffic?" returned a summary of the whole network labelled as the user's view. Both now use the area around the user when their position is known. When it is not, a traffic answer says it covers the whole network rather than presenting it as local.

The documentation assistant reads all of the documentation

The assistant on docs.navvo.io used to answer from a selection of pages, trimmed to fit its model's limit — particularly in Arabic, where the full documentation did not fit. It now reads the entire documentation in the language of the question on every answer, and repeat questions are answered faster because the unchanged documentation is reused between them. When something is not documented it says so and points to the nearest page instead of guessing, and an error now shows a short, plain message rather than technical detail.

An AI provider outage can no longer go unnoticed

For about three weeks from 28 August, Atlas quietly ran on its small local model because the hosted model it was configured to use had been retired by its provider. Every request failed over correctly, but the reason was discarded and only an HTTP status was logged, and the provider test in the admin panel still reported it as connected. The provider's own error message is now recorded (with any credential removed), each provider's successes, failures and truncated answers are tracked across every server, a provider that fails with a permanent error is set aside instead of being retried on every request, and the admin test now proves the model actually answers rather than only that the provider responds. Each request also has a single overall time limit across all providers.

The first question after an update no longer waits for the documentation to be read

The assistant reuses the documentation it has already read between questions, which is what makes most answers start within one to three seconds. After each update of the documentation site that reuse started from nothing, so the first question in each language waited around seventeen seconds while the assistant read the whole documentation. It now reads both languages once as the site starts, so the first visitor after an update typically gets an answer starting in a few seconds rather than waiting for that full read.

Route planning works again

Starting a navigation session failed with "No route could be planned between the origin and destination" for every pair of points, and address search returned nothing from the base map. Neither was a data problem: the routing and address services could not be reached at all. A server restart on 18 September replayed a saved firewall snapshot that still contained forwarding rules for an older generation of those services, pointing three internal addresses at destinations that no longer existed. Anything listening on them was invisible, however healthy it was. The services now run on their own addresses, clear of those rules, and route planning, address search, reverse lookup and the bridge and tunnel geometry behind junction views are all working again. Nothing about the public API changed — same endpoints, same responses.

Supporting services recover on their own again

The supervisor that restarts stopped services now also covers the routing engine's companion data service, which it had never watched — so when that stopped it stayed stopped for three days and junction views quietly lost their geometry. It also judges that service by asking it a real question rather than by trusting a status page the service itself answers incorrectly, waits for it to finish opening its database before deciding it is broken, and rebuilds it from scratch instead of restarting it, because a restart preserves a lock file that prevents it from ever starting cleanly. After rebuilding any of these services it also refreshes their network rules, since a leftover rule from the previous generation can silently send traffic to an address that no longer exists.