RAG Enrichment

Daimon supports transparent Retrieval-Augmented Generation (RAG): before every chat request, the sidecar automatically queries a vector store and prepends the top results as a system message. The LLM receives relevant context without any change to your client code.

---

How it works

1. A vector store is configured and loaded with documents.
2. An LLM component declares memory_store: <store-name>.
3. On each /v1/converse/{component} request, daimon:
4. Finds the last user message in the request.
5. Queries the named vector store for the top 5 most relevant documents.
6. Prepends a system message containing those documents to the conversation history.
7. Calls the LLM with the enriched history.

The client sees no difference in the API surface — it still sends messages and receives SSE chunks.

---

Configuration

Declare the store before the LLM component, then add memory_store: to the LLM component:

components:

  # 1. Embedder (if the store needs one)
  - name: embedder
    type: embedding/openai
    metadata:
      base_url: http://localhost:11434/v1
      model: nomic-embed-text
      dimensions: "768"

  # 2. Vector store
  - name: docs
    type: qdrant
    metadata:
      base_url: http://localhost:6333
      collection: product-docs
      embedder: embedder
      create_if_missing: "true"

  # 3. LLM with RAG enabled
  - name: claude
    type: anthropic
    memory_store: docs              # ← enables transparent RAG
    metadata:
      default_model: claude-opus-4-7

---

Populating the store

Use the memory HTTP API or the SDK before the first chat request:

PythonTypeScriptcurl

import daimon

client = daimon.Client()
store = client.memory("docs")

# Bulk-load documents
for doc in my_documents:
    store.upsert(doc["text"], metadata={"source": doc["url"]})

import { Client } from 'daimon-client';

const client = new Client();
const store = client.memory('docs');

for (const doc of myDocuments) {
    await store.upsert(doc.text, { metadata: { source: doc.url } });
}

curl -X POST http://127.0.0.1:3500/v1/memory/docs \
  -H "Content-Type: application/json" \
  -d '{"content": "Daimon listens on port 3500 by default.", "metadata": {"source": "docs"}}'

---

Explicit tool access

In addition to transparent RAG, every vector store automatically generates two tools the LLM can call:

• {store_name}_search — search for documents matching a query
• {store_name}_upsert — store a new document

These are injected alongside MCP tools and available in every chat request. The LLM can choose when to search rather than having context injected on every turn — useful for agentic workloads where retrieval is conditional.

# No special client code needed — the LLM decides when to call the tools
reply = client.chat("claude", "What does our documentation say about authentication?")
# The model may call docs_search("authentication") automatically

---

Choosing a store

Store	Embedding	Good for
inmemory	BM25 (lexical)	Development, keyword-heavy docs
chroma	Server-side (SentenceTransformer)	Quick setup, no embedding config
qdrant	Configurable endpoint	Production, high-volume
redis	Configurable endpoint	Existing Redis Stack infrastructure
pgvector	Configurable endpoint	Existing Postgres infrastructure

For production RAG, prefer qdrant or pgvector with nomic-embed-text (via Ollama) or text-embedding-3-small (OpenAI).

---

Failure behaviour

If the vector store query fails (e.g. the backend is down), the error is logged as a warning and the request proceeds without enrichment. The LLM receives the original message history unchanged.