Adapters Module

Knowledge base adapters for readers, vector stores, retrievers, and rerankers

Adapters Module

The Adapters module provides concrete implementations of knowledge base components including readers, vector stores, retrievers, and rerankers.

Import

from praisonai.adapters import (
    # Readers
    AutoReader, TextReader, MarkItDownReader, DirectoryReader,
    
    # Vector Stores
    ChromaVectorStore,
    
    # Retrievers
    BasicRetriever, FusionRetriever,
    
    # Rerankers
    LLMReranker
)

Quick Example

from praisonai.adapters import AutoReader, ChromaVectorStore, BasicRetriever

# Load documents
reader = AutoReader()
docs = reader.load("./documents/")

# Store in vector database
store = ChromaVectorStore(namespace="my_docs")
store.add(
    texts=[d.content for d in docs],
    embeddings=get_embeddings([d.content for d in docs]),
    metadatas=[d.metadata for d in docs]
)

# Retrieve
retriever = BasicRetriever(
    vector_store=store,
    embedding_fn=get_embedding
)
results = retriever.retrieve("search query", top_k=5)

Features

• Readers: Load documents from files, directories, URLs, and glob patterns
• Vector Stores: Store and query document embeddings (ChromaDB, Pinecone)
• Retrievers: Find relevant documents (Basic, Fusion, Recursive, AutoMerge)
• Rerankers: Improve result relevance (LLM, CrossEncoder, Cohere)

Module Structure

praisonai/adapters/
├── __init__.py          # Lazy loading exports
├── readers.py           # Document readers
├── vector_stores.py     # Vector store adapters
├── retrievers.py        # Retrieval strategies
└── rerankers.py         # Reranking implementations

Available Components

Readers

Class	Description
AutoReader	Automatic source detection and routing
TextReader	Plain text files (.txt, .log)
MarkItDownReader	Rich documents (PDF, DOCX, etc.)
DirectoryReader	Recursive directory loading
GlobReader	Glob pattern matching
URLReader	Web page content

Vector Stores

Class	Description	Requirements
ChromaVectorStore	Local persistent storage	chromadb
PineconeVectorStore	Cloud vector database	pinecone

Retrievers

Class	Description
BasicRetriever	Simple vector similarity
FusionRetriever	Multi-query with RRF
RecursiveRetriever	Depth-limited expansion
AutoMergeRetriever	Adjacent chunk merging

Rerankers

Class	Description	Requirements
LLMReranker	LLM-based scoring	OpenAI/Anthropic API
CrossEncoderReranker	Neural reranking	sentence-transformers
CohereReranker	Cohere Rerank API	cohere

Lazy Loading

All adapters use lazy loading to minimize import time:

# Only loads when accessed
from praisonai.adapters import ChromaVectorStore  # Fast import

# Actual loading happens on first use
store = ChromaVectorStore()  # chromadb loaded here

Example: Full RAG Pipeline

from praisonai.adapters import (
    AutoReader,
    ChromaVectorStore,
    FusionRetriever,
    LLMReranker
)
from praisonaiagents import Agent

# 1. Load documents
reader = AutoReader()
docs = reader.load("./knowledge_base/")

# 2. Store with embeddings
store = ChromaVectorStore(namespace="kb")
store.add(
    texts=[d.content for d in docs],
    embeddings=get_embeddings([d.content for d in docs])
)

# 3. Create retriever with fusion
agent = Agent(instructions="Query assistant")
retriever = FusionRetriever(
    vector_store=store,
    embedding_fn=get_embedding,
    llm=agent,
    num_queries=3
)

# 4. Create reranker
reranker = LLMReranker(model="gpt-4o-mini")

# 5. Query pipeline
query = "How to deploy Python apps?"
results = retriever.retrieve(query, top_k=20)
reranked = reranker.rerank(query, [r.text for r in results], top_k=5)

for r in reranked:
    print(f"Score: {r.score:.3f} - {r.text[:100]}...")

CLI Integration

The adapters power the praisonai knowledge CLI commands:

# Add documents (uses readers)
praisonai knowledge add ./docs/

# Query (uses vector store + retriever + reranker)
praisonai knowledge query "search query" \
  --vector-store chroma \
  --retrieval fusion \
  --reranker llm

Related

• Readers Module - Document loading
• Vector Store Module - Vector storage
• Retrieval Module - Document retrieval
• Reranker Module - Result reranking