Phidata

Provides a unified Python API that abstracts across OpenAI, Anthropic, Google, Ollama, and other LLM providers through a common Agent class. Internally routes requests to provider-specific SDK clients while normalizing request/response formats, enabling seamless provider switching without code changes. Handles model-specific parameter mapping (e.g., temperature, max_tokens) and response parsing across different API schemas.

Enables agents to invoke external functions through a schema-based tool registry that automatically generates OpenAI/Anthropic-compatible function schemas from Python function signatures and docstrings. The framework handles schema generation, function invocation, and response parsing, supporting both synchronous and asynchronous tool execution. Tools are registered declaratively and the agent automatically includes them in function_calling requests to the LLM.

Enables agents to use chain-of-thought reasoning patterns where the LLM explicitly breaks down problems into steps before generating final answers. The framework automatically constructs prompts that encourage step-by-step reasoning, captures intermediate reasoning steps, and uses them to improve final outputs. Supports both explicit chain-of-thought (shown to users) and implicit reasoning (internal only).

Allows customization of agent behavior through system prompts and personality configuration. Developers can define custom instructions, constraints, tone, and behavioral guidelines that shape how agents respond. System prompts are automatically prepended to all LLM calls, ensuring consistent agent behavior across interactions. Supports prompt templates with variable substitution for dynamic configuration.

Provides utilities for processing various document formats (PDF, markdown, plain text, web pages) and chunking them into manageable pieces for embedding and retrieval. Handles document parsing, text extraction, metadata preservation, and intelligent chunking strategies (semantic, fixed-size, sliding window). Chunks are automatically embedded and stored in knowledge bases for RAG.

Phidata integrates vision models (OpenAI Vision, Claude Vision, etc.) for analyzing images and providing detailed descriptions, object detection, text extraction (OCR), and visual reasoning. The framework handles image encoding, provider-specific vision API calls, and response parsing for vision-enabled agents.

Provides a pluggable memory system that stores conversation history, tool call results, and agent state across sessions. Supports multiple backends (in-memory, SQLite, PostgreSQL) and automatically manages message history, context windows, and memory summarization. Memory is attached to agents and automatically updated after each interaction, enabling stateful multi-turn conversations and long-running agent instances.

Integrates vector-based retrieval with agents through a Knowledge class that chunks documents, generates embeddings, and stores them in vector databases (Pinecone, Weaviate, Chroma, etc.). Agents can retrieve relevant documents before generating responses, augmenting their knowledge with external sources. The framework handles embedding generation, similarity search, and result ranking automatically.

Enables agents to generate structured, validated outputs by specifying Pydantic models as response schemas. The framework automatically constructs prompts that guide the LLM to produce JSON matching the schema, then parses and validates responses. Supports nested models, optional fields, and custom validators, ensuring type-safe agent outputs.

Supports building teams of specialized agents that collaborate to solve complex problems. Agents can delegate tasks to other agents, share memory and knowledge bases, and coordinate through a supervisor or hierarchical structure. The framework provides patterns for agent-to-agent communication, result aggregation, and workflow coordination without requiring manual message passing.

Enables agents to stream responses token-by-token to clients, providing real-time feedback and reducing perceived latency. The framework handles streaming protocol negotiation with different LLM providers, token buffering, and graceful error handling during streams. Supports both text streaming and structured output streaming with partial validation.

Provides built-in logging and monitoring of agent execution, capturing all LLM calls, tool invocations, memory updates, and decision points. Generates detailed execution traces that show the agent's reasoning path, including prompts sent to the LLM, responses received, and tool results. Traces can be exported for debugging, auditing, or performance analysis.

Integrates web search capabilities into agents, allowing them to retrieve current information from the internet when needed. Agents can decide when to search, formulate queries, and incorporate results into responses. Supports multiple search providers (Google, DuckDuckGo, Bing) and handles result parsing, ranking, and deduplication automatically.

Supports asynchronous execution of agents and concurrent invocation of multiple tools in parallel. Agents can await results from multiple tools simultaneously, reducing latency for workflows with independent tool calls. Built on Python's asyncio, enabling integration with async frameworks (FastAPI, aiohttp, etc.) without blocking.