Adapter guide

The generic HTTP adapter supports query parameters, JSON fields, nested request templates, Bearer authentication, wildcard response paths, retries, and streamed OpenAI-style chunks. Streaming requests are consumed incrementally and report first-token latency.

Built-in presets:

type	Defaults
langserve / langchain	POST /invoke, JSON input, answer output
llamaindex	POST /query, JSON query, answer response
dify	POST /v1/chat-messages, JSON query, answer answer
openai	POST /chat/completions, request template with messages, answer choices.0.message.content

For example, array paths can use metadata.retriever_resources.*.document_id. If an integration needs custom behavior, publish a Python package with the ragproof.adapters entry-point group:

[project.entry-points."ragproof.adapters"]
my_rag = "my_package:build_adapter"

See examples/ for LangServe, Dify, OpenAI-compatible, and the local mock server configurations.

Structured citations can be mapped explicitly when the response returns objects instead of strings:

citations_path: data.citations
citation_id_path: document.id
citation_text_path: title

When an API explicitly tells clients that it returned a deterministic fallback, make that part of the evaluation contract instead of inferring it from answer text:

fallback_path: fallback
expected_fallback: false

With this mapping, fallback: true, a missing field, or a non-boolean value becomes a response_contract error for the sample. The bundled examples/knowledgeops.yaml uses this for /ai/react/chat.

The report records first_token_latency_ms, total latency, and output character count for streamed responses. Character count is intentionally used instead of a tokenizer-specific token count so the adapter remains framework-agnostic.