Quickstart: Your First Agent

Build an AI agent with semantic memory in 7 minutes

⏱️ Time: 7 minutes 🎯 Level: Beginner 💻 SDK: All SDKs (Step 4 streaming: TypeScript only)

What You'll Build

An AI agent that:

✅ Ingests documents into semantic memory
✅ Searches by meaning (not just keywords)
✅ Answers questions with citations
✅ Streams responses in real-time (TypeScript)
✅ Calls tools to extend capabilities

Prerequisites

Required setup:

Sign up (30 seconds)
Create project (1 minute)
Configure credentials (1 minute)
✅ Verify: Run hello.ts from credentials page

For Step 4 (TypeScript streaming): OpenAI API key

Install dependencies:

npm install graphlit-client dotenv
npm install openai  # For Step 4 only

Step 1: Ingest Content

Add a document to semantic memory:

import { Graphlit } from 'graphlit-client';

const graphlit = new Graphlit();

async function main() {
  const content = await graphlit.ingestUri(
    'https://arxiv.org/pdf/1706.03762.pdf',
    'Attention Paper',
    undefined,
    undefined,
    true, // Wait for processing to complete
  );

  console.log(`✅ Document ingested: ${content.ingestUri.id}`);
}

main();

What happens: Graphlit downloads the PDF, extracts text, generates embeddings, and stores in semantic memory.

Expected output:

✅ Document ingested: 01234567-89ab-cdef-0123-456789abcdef

Python/.NET: Get this code in your language instantly:

Step 2: Search Your Memory

Query ingested content by meaning:

import { Graphlit } from 'graphlit-client';

const graphlit = new Graphlit();

async function main() {
  const results = await graphlit.queryContents({
    filter: { search: 'transformer architecture innovations' },
  });

  console.log(`Found ${results.contents.results.length} documents:`);
  
  for (const item of results.contents.results) {
    console.log(`📄 ${item.name}`);
  }
}

main();

Semantic search: Finds documents by meaning, not just keyword matching. Try searching for "attention mechanism" and see it find the transformer paper.

Expected output:

Found 1 documents:
📄 Attention Paper

Step 3: RAG Conversation

Ask questions about your content:

import { Graphlit } from 'graphlit-client';

const graphlit = new Graphlit();

async function main() {
  // Reference the document from Step 1
  const content = await graphlit.ingestUri(
    'https://arxiv.org/pdf/1706.03762.pdf',
    'Attention Paper',
    undefined,
    undefined,
    true,
  );

  // Create conversation scoped to this document
  const conversation = await graphlit.createConversation({
    name: 'Q&A Session',
    filter: { contents: [{ id: content.ingestUri.id }] }
  });

  // Ask questions
  const answer = await graphlit.promptConversation(
    'What are the key innovations in this paper?',
    conversation.createConversation.id,
  );

  console.log(answer.promptConversation.message?.message);
}

main();

What happens: Graphlit retrieves relevant sections, injects context into the LLM, and generates an answer with citations.

Expected output:

The paper introduces the Transformer architecture, which relies entirely on 
self-attention mechanisms rather than recurrence or convolutions. Key innovations 
include multi-head attention and positional encodings.

Get this code in your language:

Step 4: Real-Time Streaming (TypeScript)

TypeScript SDK only: Python and C# SDKs use synchronous promptConversation() from Step 3. Real-time streaming is TypeScript-specific.

Setup

Add to your .env:

OPENAI_API_KEY=your_openai_key

Get your key from platform.openai.com/api-keys.

Code

import { Graphlit } from 'graphlit-client';
import { OpenAI } from 'openai';
import {
  SpecificationTypes,
  ModelServiceTypes,
  OpenAiModels,
} from 'graphlit-client/dist/generated/graphql-types';

const graphlit = new Graphlit();

// Enable streaming with OpenAI client
graphlit.setOpenAIClient(new OpenAI());

async function main() {
  const spec = await graphlit.createSpecification({
    name: 'Assistant',
    type: SpecificationTypes.Completion,
    serviceType: ModelServiceTypes.OpenAi,
    openAI: {
      model: OpenAiModels.Gpt4O_128K,
      temperature: 0.7
    }
  });

  await graphlit.streamAgent(
    'Explain transformer attention in simple terms',
    (event) => {
      if (event.type === 'message_update') {
        process.stdout.write(event.message.message);
        if (!event.isStreaming) {
          console.log('\n[complete]');
        }
      }
    },
    undefined,
    { id: spec.createSpecification.id },
  );
}

main();

What happens: Tokens stream in real-time as the AI generates the response (like ChatGPT's typing effect).

Expected output:

Transformer attention is a mechanism that allows the model to focus on different 
parts of the input when processing each token. Think of it like highlighting the 
most relevant words in a sentence when trying to understand each word's meaning.
[complete]

Step 5: Add Tool Calling

Give your agent functions to call:

import { Graphlit } from 'graphlit-client';
import { OpenAI } from 'openai';
import {
  SpecificationTypes,
  ModelServiceTypes,
  OpenAiModels,
  ToolDefinitionInput,
} from 'graphlit-client/dist/generated/graphql-types';

const graphlit = new Graphlit();
graphlit.setOpenAIClient(new OpenAI());

// Define tool
const searchTool: ToolDefinitionInput = {
  name: 'search_memory',
  description: 'Search semantic memory for documents',
  schema: JSON.stringify({
    type: 'object',
    properties: {
      query: { type: 'string', description: 'Search query' },
    },
    required: ['query'],
  }),
};

// Tool implementation
const toolHandlers = {
  search_memory: async (args: { query: string }) => {
    const results = await graphlit.queryContents({
      filter: { search: args.query },
    });
    return results.contents.results.map((c) => c.name);
  },
};

async function main() {
  const spec = await graphlit.createSpecification({
    name: 'Agent with Tools',
    type: SpecificationTypes.Completion,
    serviceType: ModelServiceTypes.OpenAi,
    openAI: { model: OpenAiModels.Gpt4O_128K }
  });

  await graphlit.streamAgent(
    'Find documents about attention mechanisms',
    (event) => {
      if (event.type === 'tool_update' && event.status === 'completed') {
        console.log(`\n🔧 Called ${event.toolCall.name}`);
      } else if (event.type === 'message_update') {
        process.stdout.write(event.message.message);
        if (!event.isStreaming) {
          console.log('\n[complete]');
        }
      }
    },
    undefined,
    { id: spec.createSpecification.id },
    [searchTool],
    toolHandlers,
  );
}

main();

What happens: The agent decides when to call your function, executes it, and uses the results in its response.

Tool calling works in all SDKs: Python, TypeScript, and C# all support defining tools and handlers.

What You've Built

In 7 minutes, you created an AI agent with:

Capability

Why It Matters

Semantic memory

Ingest and search documents by meaning

RAG conversations

Q&A grounded in your content

Real-time streaming

TypeScript token-by-token responses

Agentic behavior

AI that calls functions to accomplish tasks

Data Flow Summary

Ingest Content → Semantic memory indexes files, messages, and pages
Create Specification → Pick the LLM and parameters for the agent
Create Conversation → Optionally scope retrieval with filters
promptConversation (all SDKs) or streamAgent (TypeScript) → Get responses
Tool Handlers → Agent can call functions when needed

Production Notes

Timeouts: For very large files, ingestUri(..., true) may exceed default timeouts. Consider wrapping in Promise.race with a timeout or polling via isContentDone.

Logging: Replace console.log with structured logging (Pino/Winston) in production services.

Secrets: Keep .env out of version control; use platform secret stores in deployment.

Rate limits: OpenAI streaming respects your account quotas. Handle 429 responses with retries.

Next Steps

Learn Advanced Patterns

AI Agents with Memory - Multi-agent systems, advanced tool patterns (15 min)

Knowledge Graph - Extract entities and relationships (20 min)

MCP Integration - Connect to your IDE (10 min)

Explore Sample Applications

📓 60+ Colab Notebooks - Run Python examples instantly

RAG & Conversations (15+ examples)
Ingestion & Preparation (6+ examples)
Knowledge Graph & Extraction (7+ examples)

🚀 Next.js Apps - Deploy-ready applications

Full-featured chat with streaming
Chat with knowledge graph visualization
Document extraction interface

💻 Streamlit Apps - Interactive Python UIs

Add More Capabilities

Different AI Models:

// Use Claude instead
import { ModelServiceTypes, AnthropicModels } from 'graphlit-client/dist/generated/graphql-types';

serviceType: ModelServiceTypes.Anthropic,
anthropic: {
  model: AnthropicModels.Claude_4_5Sonnet
}

Multiple Documents:

// Upload multiple PDFs
const urls = [
  'https://example.com/doc1.pdf',
  'https://example.com/doc2.pdf',
];

const ids = [];
for (const url of urls) {
  const content = await graphlit.ingestUri(url, undefined, undefined, undefined, true);
  ids.push(content.ingestUri.id);
}

// Create conversation with all documents
const conversation = await graphlit.createConversation({
  name: 'Multi-Document Chat',
  filter: { contents: ids.map(id => ({ id })) }
});

Custom Tools:

// Add a database query tool
const dbTool: ToolDefinitionInput = {
  name: 'query_database',
  description: 'Query the customer database',
  schema: JSON.stringify({
    type: 'object',
    properties: {
      query: { type: 'string', description: 'SQL query' },
    },
    required: ['query'],
  }),
};

Complete Examples

Full working code:

TypeScript SDK README - All examples tested and verified
Sample Repository - 60+ working examples
Next.js Apps - Full-stack applications

Troubleshooting

"streamAgent is not a function" (Python/C#)

Use prompt_conversation() (Python) or PromptConversation() (C#). Streaming is TypeScript-only. See Step 3 for the universal pattern.

"OpenAI API key not found"

Only needed for TypeScript streamAgent() (Step 4). Add to .env:

OPENAI_API_KEY=your_key

Get your key from platform.openai.com/api-keys.

"Content not finished processing"

Use isSynchronous: true (fifth parameter) in ingestUri() to wait for completion:

await graphlit.ingestUri(url, name, undefined, undefined, true);

"Module not found: dotenv"

Install dotenv:

npm install dotenv

Need Help?

Discord Community - Get help from the Graphlit team and community

Ask Graphlit - AI code assistant for instant SDK code examples

TypeScript SDK Docs - Complete API reference

Sample Gallery - Browse working examples

Last updated 14 days ago

Was this helpful?