ó À‘´j ãó•Srg)uüThe answering playground: ask questions of an ingested document and see each answer grounded on the page it came from. `app` is the FastAPI server (`quber playground` starts it) and owns every route. It serves the page from `static/`, lists documents and resolves cited ids through `db`, and runs uploads as subprocess jobs — `quber fuse` and `quber figure`, then `ingest_fusion`. When `gpu.configured()` is true, the extracting stage instead runs `quber table` here, sends the docling parse to the RunPod worker through `gpu.parse`, then runs `quber fuse --artifacts-dir`. Ingest embeddings go to the worker through `gpu.embed` on such a host too. The upload's track selects the model line for the whole job: the figure scan's model and the ingest's chunking stream together. `ingest_fusion` creates each document's rows in the `ade_playground` Postgres schema that `schema.sql` defines. It is not the only writer. `app` stamps upload metadata on the documents table, applies metadata edits, and deletes or replaces documents. `backfill_boxes` rewrites grounding boxes. `jobs` writes the job tables that `jobs.sql` defines in the same schema. Outside the package, the rj-reit-sweep skill's scripts write the `sweep_answers` cache table. A question flows retrieval-first: `retrieval` fuses vector and keyword ranking over the ingested chunks (`embedding` supplies the vectors), a ranker chosen by settings orders them (`selection` reads a window of the fused ranking with Haiku; `ranking` judges every page and chunk with Jev), and `answers` turns the retrieved context into one `Answer` whose payload says what shape the reply took. `agent` builds the Claude model for every agent defined in this package (`anthropic_model`, `answer_model`) and the context format the answering agents share (`build_context`). The cover-page attributor that uploads run and the footnote lookup that ingest runs are built in `quber.agents` and choose their own credential there. Its prose `answer` agent is off the request path and is the benchmark's baseline. `benchmark` measures the answering path against FinanceBench and is not imported by the server. Where the local corpus lives (staged PDFs, parse artifacts, uploads) comes from `get_settings().playground`, never from a path relative to this package. N)Ú__doc__©óÚ7/home/mande/repo/quber/src/quber/playground/__init__.pyÚrs ðò'r