RAG Pipeline

The RAG pipeline handles document upload, chunking, vector indexing, retrieval, and context injection before model inference. Eridian manages the full lifecycle - your application sends a query; the orchestration layer retrieves relevant chunks and assembles context within token budgets automatically.

Document Upload

POST /v1/rag/documents
Authorization: Bearer eridian_sk_...
Content-Type: multipart/form-data

Upload PDF, DOCX, or plain text documents. Documents are chunked, embedded, and stored in an encrypted vector store scoped to your project. Processing status is available via GET /v1/rag/documents/{id}.

Processing States

Chunking Strategies

Legal workflows often pair legal_clause chunking with similarity_threshold: 0.78 because clause wording varies more than FAQ-style content.

Retrieval Configuration

{
  "rag_config": {
    "top_k": 5,
    "similarity_threshold": 0.75,
    "max_context_tokens": 4096,
    "rerank": true,
    "collection_ids": ["col_msa_templates"]
  }
}

Enable RAG on inference requests with "features": ["rag"]. Responses include:

• eridian.rag_injected - whether context was added
• eridian.rag_chunks - count of chunks injected
• eridian.rag_document_ids - source documents referenced

Context Injection

Retrieved chunks are injected into the prompt context before inference using a stable template:

<context> [chunk 1 metadata + text] [chunk 2 metadata + text] </context> <user_query> ...

The observability dashboard shows which chunks were used for each request, supporting audit and debugging in regulated workflows. Cache keys include document version hashes so corpus updates invalidate stale cached answers.

Delete and Reindex

DELETE /v1/rag/documents/{id}
POST /v1/rag/collections/{id}/reindex

Deletion is soft by default with 30-day retention for compliance holds. Reindex rebuilds embeddings after chunking strategy changes.

Error Codes

See Inference for RAG feature flags and Semantic Caching for cache interaction with RAG prompts.