conversekit 0.0.1

A modular Python library for building conversational AI backends.

🚧 Status: ConverseKit is currently in Pre-Alpha. The package has been published to reserve the project name while the architecture and implementation are actively being developed.

ConverseKit

Build conversational AI backends by composing infrastructure instead of implementing it.

An open-source Python library for building modular conversational AI applications.

---

Why ConverseKit?

Every conversational AI application ends up implementing the same backend infrastructure:

• Conversation management
• Document ingestion
• Parsing & chunking
• Embedding generation
• Vector database integration
• Retrieval pipelines
• Memory management
• Context assembly
• Prompt orchestration
• LLM integrations

Most of this code isn't unique to your application.

ConverseKit eliminates this repetitive engineering by providing reusable, configurable modules that work together through a consistent architecture.

Instead of writing backend infrastructure, developers simply compose it.

---

Philosophy

ConverseKit is built around one simple idea:

Developers should build AI applications, not AI infrastructure.

Every major component is modular, interchangeable, and independently configurable.

Choose the techniques you want.

ConverseKit handles how they work together.

---

Features

📄 Modular Document Pipeline

Configure your document pipeline instead of implementing it.

• Multiple document loaders
• Configurable parsers
• Chunking strategies
• Metadata extraction
• Embedding generation
• Vector database indexing

---

🔍 Pluggable Retrieval

Swap retrieval strategies without modifying your application.

Examples include:

• Naive RAG
• Hybrid Search
• Parent Document Retrieval
• Multi Query Retrieval
• Corrective RAG
• Custom Retrieval Strategies

---

🧠 Modular Memory

Support different memory systems through a common interface.

Examples:

• Sliding Window Memory
• Long-Term Memory
• Branch Memory
• Custom Memory Providers

---

🌳 Branch-aware Conversations

ConverseKit treats conversations as first-class objects.

Applications can support:

• Multiple conversations
• Conversation branching
• Branch switching
• Branch-aware context assembly
• Context inheritance

---

🧩 Context Assembly

Context assembly is a core component of ConverseKit.

Instead of simply retrieving documents, ConverseKit intelligently combines:

• User message
• Conversation history
• Retrieved documents
• Long-term memory
• Branch history
• System prompts

before generating a response.

---

📊 Evaluation

Compare different pipelines scientifically.

Measure:

• Context Relevance
• Faithfulness
• Answer Relevance
• Precision@K
• Recall@K

Developers can evaluate retrieval strategies and choose the best-performing pipeline for their application.

---

🔌 Provider Abstraction

Swap providers without changing application logic.

Supported provider types include:

• LLM Providers
• Embedding Providers
• Vector Databases
• Rerankers

---

Example

from conversekit import Application

app = Application(

    ingestion=DocumentPipeline(

        parser=PyMuPDFParser(),

        chunker=RecursiveChunker(),

        embedding=BGEEmbedding(),

        vectordb=Qdrant()
    ),

    retrieval=HybridRetriever(),

    memory=SlidingMemory(),

    llm=GeminiProvider()

)

response = app.chat(
    "Summarize Chapter 5"
)

Notice that no backend infrastructure is implemented.

The developer simply describes the pipeline they want.

---

Modular Architecture

                User
                  │
                  ▼
          Conversation Manager
                  │
                  ▼
        Document Retrieval Pipeline
                  │
                  ▼
          Context Assembly
                  │
                  ▼
          Prompt Builder
                  │
                  ▼
             LLM Provider
                  │
                  ▼
              Response

Every stage is independently replaceable.

---

Design Principles

• Modular by default
• Interface-driven architecture
• Provider independent
• Retrieval independent
• Memory independent
• Easy to extend
• Easy to understand
• Easy to evaluate

---

Example Applications

ConverseKit can serve as the backend for:

• Educational Tutors
• PDF Chat Applications
• Legal Assistants
• Medical Assistants
• Customer Support Bots
• Research Assistants
• Enterprise Knowledge Assistants
• Internal Company Chatbots

---

Roadmap

Core

• Conversation Management
• Document Ingestion
• Context Assembly
• Memory Framework
• Retrieval Framework
• Prompt Builder
• Provider Abstractions

Retrieval

• Naive RAG
• Hybrid RAG
• Parent Document Retrieval
• Corrective RAG
• Multi Query Retrieval
• Branch-aware Retrieval

Evaluation

• Retrieval Benchmarks
• Faithfulness Metrics
• Context Metrics
• Latency Benchmarks

Future

• Agentic Workflows
• Vision Support
• MCP Integration
• Multimodal Pipelines

---

Current Status

🚧 ConverseKit is currently under active development.

The project is being designed with a strong focus on modularity, extensibility, and developer experience.

---

Vision

The goal of ConverseKit is to become the foundation developers reach for whenever they build a conversational AI application.

Whether you're creating a prototype, a research project, a startup MVP, or a production system, ConverseKit should allow you to assemble a complete conversational AI backend by configuring reusable components instead of rebuilding infrastructure from scratch.

---

License

MIT License