RAG Pipelines

Retrieval-Augmented Generation with 5 built-in retrieval strategies.

AgentOven includes a built-in Retrieval-Augmented Generation (RAG) pipeline with 5 retrieval strategies. Ingest documents, chunk them, generate embeddings, store vectors, and retrieve context for LLM generation — all through a unified API.

Retrieval Strategies

Choose the right strategy based on your use case. Each strategy optimizes for different trade-offs between speed, accuracy, and cost.

Naive

Simple embed → search → generate pipeline. Best for straightforward Q&A where the query closely matches document content.

Query → Embed → Vector Search → Top-K → LLM Generate

Sentence Window

Retrieves matching sentences then expands context by including surrounding sentences. Provides focused yet contextualized results.

Query → Embed → Search Sentences → Expand Window → LLM Generate

Parent Document

Searches at the chunk level but returns full parent documents. Great when small chunks match but the full document context is needed.

Query → Embed → Search Chunks → Return Parents → LLM Generate

HyDE

Hypothetical Document Embeddings — first generates a hypothetical answer using the LLM, then embeds and searches for real matches. Excels when queries are abstract.

Query → LLM Hypothesize → Embed Hypothesis → Search → LLM Generate

Agentic

LLM decides whether to search, refine the query, or synthesize from existing context. Most intelligent but highest latency and cost.

Query → LLM Decide → [Search | Refine | Synthesize] → Iterate → LLM Generate

Document Ingestion

The ingestion engine handles the full pipeline from raw documents to searchable vectors:

1. Upload documents via API
2. Text is split into chunks (configurable size, overlap, separator)
3. Chunks are embedded in batches using the configured embedding driver
4. Vectors are upserted into the configured vector store
5. Original text and metadata preserved for retrieval

Usage Example

Ingest documents

$ curl -X POST http://localhost:8080/api/v1/rag/ingest \
  -H "Content-Type: application/json" \
  -H "X-Kitchen: default" \
  -d '{
    "documents": [
      {
        "id": "doc-1",
        "content": "AgentOven is an open-source agent control plane...",
        "metadata": { "source": "readme", "version": "0.2" }
      }
    ],
    "chunk_size": 512,
    "chunk_overlap": 50
  }'

Query with RAG

$ curl -X POST http://localhost:8080/api/v1/rag/query \
  -H "Content-Type: application/json" \
  -H "X-Kitchen: default" \
  -d '{
    "question": "What is AgentOven?",
    "strategy": "naive",
    "top_k": 5,
    "model": "gpt-4o-mini"
  }'

Chunker Configuration

The text chunker is configurable to optimize for your document structure:

{
  "chunk_size": 512,
  "chunk_overlap": 50,
  "separator": "\n\n"
}

• chunk_size — Maximum characters per chunk
• chunk_overlap — Overlap between adjacent chunks to preserve context
• separator — Primary split boundary (falls back to sentence/word splitting)

API Endpoints