Spaces:
Paused
Paused
Commit History
-
Actually sync src/ from aip_chatbot@caebbcd (previous sync was docs-only)
The prior "Sync from aip_chatbot@caebbcd" commit (a3d6262) only carried over the entrypoint.sh comment update. It never copied the actual source change: this Space was still serving the old planning_server.py with the live Wikipedia lookup (aircraft_type_specs / typespec.py) registered and callable, which is exactly why the deployed agent kept citing Wikipedia for aircraft variants after the tool was supposedly removed. This commit does the sync that was skipped: - Remove aircraft/typespec.py and the aircraft_type_specs tool from planning_server.py. - Add ga_catalog_import.py and knowledge_import.py (GA catalog + Wikidata/ DBpedia sources) and the bundled ga_aircraft_catalog.xlsx. - Update database.py/__init__.py/build_cli.py for the four-source model. Also fix .gitignore: its blanket `data/` rule was anchoring to every `data/` directory in the tree, including src/vfrplan/aircraft/data/, which silently kept the bundled GA catalog xlsx out of every commit including this one until …aa43d95
-
Start serving before the ingest, not after it
The last deploy died with "Launch timed out, workload was not healthy after 30 min". The entrypoint ran the whole ingest ahead of exec, so the port stayed shut until it finished, and the newly added visual chart fetches -- ~126 PDFs on top of the AD 2 and Guia VFR pages, all at the deliberate 1 request/second -- pushed a cold Spanish ingest past Hugging Face's launch deadline. Everything was downloaded and nothing was served. Backgrounding the data preparation is what DEPLOY.md already describes: a server with no snapshot "starts, answers, and reports honestly that it has no data", and snapshot_status reports whatever is loaded at the moment it is asked, so a client arriving mid-ingest gets a truthful partial answer instead of a refused connection. Reads during that window are safe now that the store opens SQLite in WAL mode with a busy timeout (aip_chatbot@7db0710), synced here in the same commit -- under the previous rollback journal a reader died with "database is locked" once the…4353c61 -
Sync from aip_chatbot@b245d0f: VAC chart text, two-code aerodromes, PDF AD 2
Brings in the ingest changes that make visual procedures searchable and add the aerodromes discovery was silently dropping: - VAC charts are extracted as text and filed as ordinary sections (chart_vac_1, ...), so reporting points and arrival, departure and communications-failure procedures are searchable. AD 2.20 and AD 2.22 routinely just say "see VAC", so these existed nowhere before. - Aerodromes holding two ICAO codes (LECU/LEVS, LEPA/LESJ) are now discovered and answer to either code. - The seven aerodromes ENAIRE publishes as a PDF rather than a web page are ingested instead of skipped. - A two-digit WEF year is no longer read as next century, which had Gibraltar's 1996 document reporting as current. NOTE for whoever deploys this: entrypoint.sh skips the ingest whenever a snapshot already exists on the attached bucket, so this rebuild alone will keep serving the old snapshot — same aerodromes, no chart sections. Set VFRPLAN_FORCE_REINGEST=1 and r…9ef3ad8
-
Sync from aip_chatbot@5229c1f
Routes aircraft-type identification through aircraft_database before aircraft_type_specs (PROMPT.md), and fixes the EUROCONTROL scraper swallowing the whole ATC-PFDB aircraft picker as the aircraft name (eurocontrol_import.py) — see aip_chatbot@5229c1f and @a678fb2 for the full explanation.2929f10
-
Sync from aip_chatbot@2fb88e6: add aircraft_database tool and free-text date/time parsing
Brings in the local FAA/EUROCONTROL aircraft performance database (new aircraft_database MCP tool, 17 tools total) and the earlier free-text date/time acceptance in plan_route/fuel_plan, neither of which had been synced to this Space before. Also: build the FAA database automatically on first boot if none exists yet (entrypoint.sh), add VFRPLAN_AIRCRAFT_DB to the Dockerfile, and update DEPLOY_HF.md/PROMPT.md accordingly. The crodri/vfrplanner bucket already has a pre-built FAA+EUROCONTROL merge uploaded via `hf cp`, so the boot-time build step will find it and skip.05ab6c1
-
Despliegue inicial de vfrplan MCP
7c25f4f