Context documents and datasets
Attach your own material and make it available to research agents.
Attach files to research
Context documents come from the research's workspace Library. Organization research uses organization files, available to members; personal research uses the user's personal files. The document picker can also upload a new file to that Library and shows the plan's storage meter.
The latest version of a research dataset is attached automatically. Uploading a new dataset version replaces the previous attached version.
Processing states
Documents can be pending, processing, ready, failed, or unsupported. Processing extracts text, creates passages, and builds a search index for that research.
Supported extraction includes text-layer PDF, DOCX, CSV, TSV, JSON, notebooks, code, and text. Scanned PDFs without a text layer and images require OCR and are marked unsupported. Corrupt files fail and can be retried. The processor limits a document to 600 passages and 2 M characters, with up to 100 documents per research.
What agents can read
Agents search only the attached passages of their current research through search_research_docs. The planner and agent history include file names, processing states, and short extracts. Retrieved passages are fenced as untrusted content.
Detaching a document or deleting its source media removes the corresponding chunks and embeddings.
Use a document in a run
A run can mount up to eight documents. Inputs are fetched through time-limited URLs and are included in the compute cache key.
import lasques
path = lasques.input_path("data.csv")The input name must match an attached document selected for the run. Inputs are downloaded into the run's inputs/ directory; they are not inlined into the code snapshot.