Skip to content

Migrating to Adapta

Kind: 📖 Reference. For teams coming from a RAG-only stack (Flowise, Cognita, R2R, or similar) or from a setup where a platform trained adapters for you. Checked 2026-10-10.

Flowise (Apache-2.0) was archived on 2026-08-13 and Cognita (Apache-2.0) on 2026-03-13. Adapta does not import their exports: you re-upload the source documents and re-create the endpoint. The documents are the asset; the pipeline around them is a few API calls.

From Flowise, Cognita or R2R

There In Adapta
A flow / chatflow / pipeline A project of type rag, with one endpoint
Document store, data source, collection The project's files (PDF, DOCX, TXT, MD, HTML), indexed by the worker
Chain or retrieval node Built in: hybrid retrieval (vector + Postgres full-text, fused), cited answers
Embedding model setting One instance-wide embedder (ADAPTA_EMBEDDING_MODEL, multilingual by default)
Prompt template The project's system prompt (system_prompt on the project)
Deployed chatflow API POST /v1/chat/completions, OpenAI-compatible, with an adp_ key bound to one endpoint

No equivalent: the visual canvas, agents and tool-calling flows, custom nodes, connectors that sync from Notion/Drive/web crawlers (upload files instead), and SSO. Adapta is a model-and-knowledge server with an admin console, not a flow builder.

A ten-line path (the stack is up per the README quick start):

API=http://localhost:8000/v1
TOKEN=$(curl -s $API/auth/login -H 'Content-Type: application/json' \
  -d '{"email":"admin@example.com","password":"admin12345"}' | jq -r .access_token)
TEAM=$(curl -s -H "Authorization: Bearer $TOKEN" $API/auth/me | jq -r '.teams[0].id')
PROJECT=$(curl -s -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' $API/projects \
  -d '{"name":"support-kb","type":"rag","base_model":"qwen2.5-1.5b-instruct","team_id":"'$TEAM'"}' | jq -r .id)
for f in docs/*.pdf; do curl -s -H "Authorization: Bearer $TOKEN" -F file=@"$f" $API/projects/$PROJECT/files; done
curl -s -H "Authorization: Bearer $TOKEN" -X POST $API/projects/$PROJECT/endpoint       # slug in the response
KEY=$(curl -s -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{"name":"app"}' $API/projects/$PROJECT/keys | jq -r .key)                          # shown once

Uploads return 202 and index on the worker; wait for the files to show as indexed (GET /v1/projects/$PROJECT/files) before asking questions. Then point your app at the endpoint exactly as in Consuming the API. Scanned PDFs and images go through OCR (about 1.3 s per page on CPU), and citations carry page and OCR provenance.

Your clients keep working if they speak the OpenAI API: change base_url and api_key, and pass the endpoint slug as model. If you relied on a flow's output shape, note that Adapta returns plain chat completions plus citations, and abstains deterministically ("I can't find that in the documents.") when retrieval finds nothing above the threshold.

From a setup where something trained adapters for you

Adapta does not train. If your platform fine-tuned a LoRA, export the adapter and bring it:

  1. Export the adapter in PEFT form (a .zip with adapter_config.json and the weights) or as a GGUF LoRA. It must target a base model in Adapta's catalog (see the import guide for the accepted aliases and chat template).
  2. Put together an eval set (classify or extract are the reliable tasks) in the guide's JSONL format.
  3. Run the same four calls as the guide's step by step: upload the eval set, upload the adapter, wait for passed or failed, then POST .../endpoint.
  4. Swap or roll back later with POST .../endpoint/promote; every change is recorded in GET .../endpoint/history.

The gate compares your adapter with the base model on your eval set, on CPU. In the recorded classify run the passing adapter scored 0.700 against the base's 0.333; an adapter that does not beat the base is refused and never serves. If you also have documents, index them in the same project and the endpoint answers from them in the adapter's voice (knowledge + behavior).

If your old stack served the adapter through Ollama, note that exporting a gated adapter as an Ollama Modelfile bundle is not built (roadmap block E2.2). Adapta can use Ollama as an optional backend for adapter-less RAG and base-model endpoints only (Operations §9).

What to expect

The stack is CPU-only and a single serial worker indexes your uploads in order, so a very large first import queues. Retrieval quality on your own documents is yours to measure: the 0.978 recall@5 figure is for the repository's 45-question golden set, not for your corpus.