Building an Enterprise Semantic Recommendation Engine with LangGraph, ChromaDB, and Multi-Agent RAG

In the enterprise landscape, recommendation engines have evolved beyond simple collaborative filtering ("users who bought X also bought Y"). Modern businesses require semantic understanding, contextual memory, and auditability. A user asking for "software that helps me manage remote team burnout" shouldn't just get HR tools; they should get a curated list based on their company’s existing tech stack, budget constraints, and previous interactions.

This article demonstrates how to build a production-grade Semantic Recommendation Agent using:

• LangGraph: For stateful, cyclic multi-agent orchestration.

• ChromaDB: For high-performance semantic vector storage.

• Enterprise RAG: Retrieval Augmented Generation with metadata filtering and re-ranking.

• Persistent Memory: To maintain user preferences across sessions.

The Real-Time Use Case: Internal Tech Stack Advisor

Scenario: An enterprise IT procurement portal where employees ask natural language questions to find approved software vendors.
The Challenge: The catalog contains 10,000+ SaaS products with dense technical documentation. Simple keyword search fails on nuanced queries like "secure alternative to Slack for healthcare compliance."
The Solution: A multi-agent system that retrieves candidates semantically, validates them against enterprise policy (RAG), remembers user department constraints, and synthesizes a ranked recommendation.

Architecture Overview

We will implement a Supervisor Multi-Agent Pattern:

1. State Manager: Holds the conversation history, retrieved docs, and user profile.

2. Retriever Agent: Queries ChromaDB with semantic similarity + metadata filters.

3. Policy Validator Agent: Checks retrieved items against internal compliance docs.

4. Recommendation Synthesizer: Generates the final response with citations.

5. Memory Writer: Updates long-term user preferences in ChromaDB.

Step-by-Step Implementation

Prerequisites

pip install langgraph langchain-chroma chromadb langchain-openai pydantic

Step 1: Define the Enterprise State

In LangGraph, state is the single source of truth. We define a structured schema to ensure type safety and traceability.

from typing import Annotated, List, Dict, Any
from typing_extensions import TypedDict
from langgraph.graph.message import add_messages
from langchain_core.documents import Document

class RecState(TypedDict):
    messages: Annotated[List[Any], add_messages]
    user_id: str
    department: str
    retrieved_docs: List[Document]
    validated_tools: List[Dict[str, Any]]
    final_recommendation: str
    iteration_count: int

Step 2: Initialize ChromaDB with Enterprise Metadata

Semantic similarity alone isn't enough. We use Chroma's metadata filtering to enforce enterprise boundaries before the LLM sees the data.

import chromadb
from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings

# In production, use PersistentClient or HTTP connection
client = chromadb.PersistentClient(path="./chroma_enterprise")

vectorstore = Chroma(
    client=client,
    collection_name="saas_catalog",
    embedding_function=OpenAIEmbeddings(model="text-embedding-3-small"),
)

# Example: Adding documents with rich enterprise metadata
# vectorstore.add_documents(
#     documents=[...],
#     metadatas=[{
#         "vendor": "AcmeCorp", 
#         "compliance": ["HIPAA", "SOC2"], 
#         "approved_departments": ["Engineering", "HR"],
#         "price_tier": "enterprise"
#     }]
# )

Step 3: Build the Specialized Agents

A. The Semantic Retriever Node

This node performs hybrid search: semantic similarity filtered by the user's department.

def retriever_node(state: RecState):
    """Retrieve relevant tools based on semantic similarity + dept filter."""
    query = state["messages"][-1].content
    
    # Enterprise-grade retrieval: Filter BEFORE semantic search
    docs = vectorstore.similarity_search(
        query,
        k=10,
        filter={"approved_departments": {"$in": [state["department"], "All"]}}
    )
    
    return {"retrieved_docs": docs}

B. The Policy Validator Node

An LLM-based gatekeeper that ensures recommendations don't violate security policies.

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)

validation_prompt = ChatPromptTemplate.from_messages([
    ("system", """You are an IT Compliance Officer. 
    Review the retrieved tools against these rules:
    1. Must have SOC2 if handling customer data.
    2. Cannot be on the banned vendor list: {banned_vendors}.
    Return ONLY valid tool names as a JSON list."""),
    ("human", "Query: {query}\nRetrieved Tools: {docs}")
])

def validator_node(state: RecState):
    chain = validation_prompt | llm
    result = chain.invoke({
        "query": state["messages"][-1].content,
        "docs": [d.metadata.get("vendor") for d in state["retrieved_docs"]],
        "banned_vendors": ["ShadowIT_Corp", "UnsecureApp"]
    })
    # Parse LLM output into structured validated_tools
    return {"validated_tools": result.content, "iteration_count": state.get("iteration_count", 0) + 1}

C. The Synthesizer & Memory Writer

Generates the answer and extracts implicit preferences for future queries.

synth_prompt = ChatPromptTemplate.from_messages([
    ("system", "You are an Enterprise Tech Advisor. Recommend tools from this validated list: {tools}. Cite sources."),
    ("human", "{query}")
])

def synthesizer_node(state: RecState):
    chain = synth_prompt | llm
    response = chain.invoke({
        "tools": state["validated_tools"],
        "query": state["messages"][-1].content
    })
    return {"final_recommendation": response.content}

Step 4: Orchestrate with LangGraph

Now we wire the nodes into a stateful graph with conditional edges.

from langgraph.graph import StateGraph, START, END

workflow = StateGraph(RecState)

# Add nodes
workflow.add_node("retriever", retriever_node)
workflow.add_node("validator", validator_node)
workflow.add_node("synthesizer", synthesizer_node)

# Define flow
workflow.add_edge(START, "retriever")
workflow.add_edge("retriever", "validator")

# Conditional edge: Re-retrieve if no valid tools found (self-healing RAG)
def should_continue(state: RecState):
    if not state["validated_tools"] and state["iteration_count"] < 3:
        return "retriever"  # Try broader search
    return "synthesizer"

workflow.add_conditional_edges("validator", should_continue)
workflow.add_edge("synthesizer", END)

# Compile with checkpointer for persistent memory
from langgraph.checkpoint.memory import MemorySaver
checkpointer = MemorySaver()
app = workflow.compile(checkpointer=checkpointer)

Step 5: Running the Recommendation Engine

config = {"configurable": {"thread_id": "user_123_dept_eng"}}

result = app.invoke(
    {
        "messages": [("human", "Need a HIPAA-compliant project management tool")],
        "user_id": "user_123",
        "department": "Engineering",
        "retrieved_docs": [],
        "validated_tools": [],
        "final_recommendation": "",
        "iteration_count": 0
    },
    config=config
)

print(result["final_recommendation"])

Why This Architecture Wins in Enterprise

Feature	Traditional RAG	LangGraph Multi-Agent + Chroma
Relevance	Pure semantic similarity	Semantic + Metadata Pre-filtering
Safety	Post-hoc guardrails	Dedicated Validator Agent in-loop
Resilience	Fails on bad retrieval	Self-healing via conditional edges
Memory	Stateless or simple buffer	Persistent state + preference extraction
Auditability	Black box	Full state trace per thread

Production Considerations

1. Chroma Scaling: For >1M documents, deploy Chroma in cluster mode with HNSW indexing tuned for your recall/latency SLA.

2. Embedding Model: Use domain-fine-tuned embeddings (e.g., bge-m3 or fine-tuned text-embedding-3) on your internal SaaS catalog for better semantic alignment.

3. Observability: Integrate LangSmith to trace each agent’s reasoning, token usage, and retrieval quality.

4. Memory TTL: Implement TTL on Chroma user-preference collections to comply with GDPR/data retention policies.

5. Async Execution: Wrap nodes in async functions and use app.astream() for real-time streaming responses in web UIs.

Conclusion

By combining ChromaDB’s metadata-aware semantic search with LangGraph’s stateful multi-agent orchestration, you move beyond toy demos to build recommendation engines that are safe, context-aware, and genuinely useful in enterprise environments. The key insight is treating recommendation not as a single retrieval step, but as a collaborative agent workflow where each specialist handles a distinct aspect of the enterprise decision-making process. This architecture is extensible: add a Pricing Negotiator agent, integrate with Jira for ticket creation, or connect to Okta for SSO validation—all within the same stateful graph.