Skip to documentation
eli.aidocs

Intake

Bulk import & onboarding

The fastest path from "empty workspace" to "grounded answers" is getting your existing material in. eli.ai supports multi-file bulk import that fans every document into the same ingestion pipeline a single save uses, and a guided first-run welcome flow that walks a new workspace through import, extraction, and its first question.

Multi-file import

Each imported file flows through the standard ingestion pipeline.

Choose or drop up to 1,000 .md, .markdown, or .txt files (5 MiB UTF-8 maximum each). The browser reads them with File.text() and posts JSON; each becomes a document through saveDocument() — the exact entry point the editor uses — so import is never a second-class path. The synchronous phase (chunking, Postgres FTS, wikilink resolution, deterministic co-occurrence edges) runs on save; the expensive phases (embeddings, LLM extraction) are enqueued to the worker, so a large import returns quickly and the graph fills in progressively.

  • Per-file isolation.One malformed file fails on its own row; the rest of the batch still imports. Each file's status is reported back individually.
  • Progressive readiness. Documents are searchable via FTS the moment they save; vector recall and the typed graph light up as the worker drains the queue.
  • Same guardrails. Imported content counts against workspace limits and flows through the same tenancy and provenance rules — nothing about bulk import bypasses row-level security.

Text transport only

Multipart upload, zip expansion, PDF/Office parsing, and attachment shadow chunks are not part of this import path. Use a content connector that supplies normalized text, or install and validate an optional connector extractor, before claiming binary-format support.

Review extracted identities

Interview transcripts import as ordinary text. Because short speaker labels can create ambiguous entity identities, inspect the duplicate-candidate inbox after extraction and merge only with source evidence. The importer does not maintain a separate speaker disambiguation queue.

The guided welcome flow

A brand-new workspace is not empty and inert — the first-run flow gives it a spine. It orients a new user around the loop that makes eli.ai useful and hands off to the right surface at each step.

  1. Bring in material

    The welcome screen leads straight to bulk import (or the editor for a first note), so the workspace has content to reason over within minutes.

  2. Watch the graph build

    As extraction runs, the flow points at the entity graph forming — entities, relations, and provenance — so the value of ingestion is visible, not abstract.

  3. Tune the ontology (optional)

    If the domain needs its own vocabulary, the flow links to ontology authoring before extraction has processed everything, so new types apply broadly.

  4. Ask the first grounded question

    Finally it drops the user into chat with a suggested question, producing an answer with clickable [Sn] citations — the payoff that closes the onboarding loop.

Seeded, not blank

Every workspace starts with a default ontology and a working retrieval pipeline, so the very first import is immediately searchable and graphed. Onboarding is about direction, not setup — there is no configuration you must complete before the product works.

Related