RAG API · Private beta

A RAG API that returns evidence, not generated answers.

Send a raw query and permitted filters. CorpusMesh returns ranked passages, source versions, citations, and an explicit answer or abstain decision. It does not generate the final answer or legal advice.

Inspect the RAG API Request developer access

POST /v1/retrieveapplication/json

Request

{
  "knowledge_base": "eu-ai-act-reference",
  "query": "How do Article 6 and Annex III work together to identify high-risk AI systems?",
  "filters": {
    "language": [
      "en"
    ],
    "jurisdiction": [
      "eu"
    ]
  },
  "top_k": 5,
  "profile": "active"
}

Illustrative response excerpt · 200

{
  "version": "2026-08-21.1",
  "results": [
    {
      "text": "2. In addition to the high-risk AI systems referred to in paragraph 1, AI systems refe…",
      "citation": {
        "label": "Article 6 Classification rules for high-risk AI systems",
        "source_anchor": "https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng#art_6",
        "source_snapshot_id": "<source-snapshot-uuid>"
      },
      "metadata": {
        "language": "en",
        "jurisdiction": "eu",
        "verification_status": "verified"
      }
    }
  ],
  "abstention": {
    "decision": "answer"
  },
  "trace_id": "ret_<trace-id>"
}

The request follows the prepared beta plan. The response excerpt illustrates its shape using the reviewed public trace; identifiers are placeholders and this is not a new live request.

MethodPOST /v1/retrieve

OutputRanked passages · JSON

Decisionanswer · abstain

Access Private beta

Contract status

This reference documents the verified private-beta runtime. It does not provide public credentials or general availability. Access is reviewed for one approved organization and one named application. Start with the quickstart, then read authentication and errors and limits before integration.

Request access

Request

The caller controls the question and requested narrowing.

Unknown fields are rejected. Filters can narrow an entitlement, never expand it. The service plan applies the final top_k cap.

Field	Type	Presence	Behavior
`query`	string	Required	Raw query text, up to 4,000 characters.
`knowledge_base`	string	Required	The entitled knowledge-base slug.
`filters.language`	string[]	Optional	Requested languages within the entitlement.
`filters.jurisdiction`	string[]	Optional	Requested jurisdictions within the entitlement.
`filters.subject`	string[]	Optional	Knowledge-base-specific subject slugs.
`top_k`	integer	Optional	1 to 100 in the schema, capped by the service plan.
`profile`	active \| UUID	Optional	Use active or an explicitly entitled retrieval-profile ID.

: "${CORPUSMESH_API_KEY:?Set CORPUSMESH_API_KEY first}"

curl --silent --show-error --fail-with-body https://corpusmesh.com/v1/retrieve \
  --request POST \
  --header "Authorization: Bearer $CORPUSMESH_API_KEY" \
  --header "Content-Type: application/json" \
  --header "Idempotency-Key: request-article-6-001" \
  --data '{
  "knowledge_base": "eu-ai-act-reference",
  "query": "How do Article 6 and Annex III work together to identify high-risk AI systems?",
  "filters": {
    "language": [
      "en"
    ],
    "jurisdiction": [
      "eu"
    ]
  },
  "top_k": 5,
  "profile": "active"
}'

Enforced boundary

The serving state is resolved before embedding the query.

The API verifies the organization, client, entitlement, immutable version, active retrieval profile, projection, and mandatory filters. A forbidden widening fails closed.

Caller sendsquery · filters · top_k

CorpusMesh pinsentitlement · version · profile · authorization

API returnspassages · citations · decision

Language, jurisdiction, subject, and profile requests are intersected with the entitlement. The API does not silently substitute broader access.

Response fields

Every passage carries its serving state and source identity.

The response identifies the corpus version, retrieval profile, trace, source document, revision, source snapshot, canonical URL, citation anchor, language, jurisdiction, and verification status.

Serving state

version, retrieval_profile, and trace_id make the retrieval state inspectable.

Citation and scope

citation.label, source_anchor, language, and jurisdiction travel with each passage.

Source lineage

source_snapshot_id identifies the captured source. Its checksums and lineage are inspectable in the public Source Trace.

Retrieval decision

An abstention is a valid response, not a hidden failure.

200 · answer

Evidence met the configured threshold.

Use the returned passages and citations as context for your own application. CorpusMesh does not generate the final answer.

200 · abstain

The query is not supported by the retrieved evidence.

results is empty and the response explains the abstention signal. Do not retry unchanged or treat it as evidence.

Evidence before integration

Inspect the retrieval result and its measured behavior.

The public artifacts show the citation path, source checksums, corpus version, evaluation fixture, and runtime-observed scores for the reviewed v1 reference. The benchmark retrieved ten passages and the public trace displays five; that evaluation is separate from the beta plan’s five-passage response cap.

Source TraceFollow passages back to source snapshots and checksums.Open the reviewed trace →RAG evaluationReview recall, ranking, abstention, and boundary results.Open the benchmark →

Plan for integration

Match the request to the approved beta boundary.

The prepared EU AI Act beta covers one English knowledge base for one named application. The service plan attached to your entitlement is authoritative.

Maximum top_k: 5
Passage length: 4,000 characters
Response size: 50,000 bytes
Request rate: 10 per minute
Concurrency: 2 requests
Daily allowance: 250 requests
Monthly allowance: 5,000 requests

Review the fixed-term beta offer and error and retry behavior. When a source changes, a candidate version must be reviewed and evaluated before activation. See how source maintenance is bounded.

Your acceptance test

Bring a workflow that can be checked.

Use these questions to prepare a technical review and define success before access is issued.

Which authorities, jurisdictions and source versions must your application cite?
Which representative questions should return evidence, and which should abstain?
What source-change cadence and review delay can your workflow accept?
What request volume, concurrency and result depth does your integration need?
Who will judge the retrieved passages and the answers your own model produces?

Use the benchmark acceptance worksheet to record these decisions. A published benchmark does not replace evaluation on your workflow.

Private beta

Evaluate the contract against your knowledge workflow.

Describe the sources, jurisdictions, update obligations, and request volume your application needs.

Request developer access