ChemCrow

ChemCrow uses a ChatZeroShotAgent pattern that interprets chemistry queries through an LLM (GPT-4 by default) to dynamically select and sequence appropriate tools from its chemistry toolkit. The agent maintains an iterative loop where tool outputs are fed back to the LLM for reasoning, enabling multi-step problem solving up to a configurable max_iterations (default 40). This differs from static tool routing by allowing the LLM to make context-aware decisions about which tools to invoke based on intermediate results.

ChemCrow wraps RDKit (a cheminformatics library) through LangChain BaseTool subclasses to enable molecular analysis without direct RDKit code. Tools parse SMILES/IUPAC inputs, compute molecular descriptors (molecular weight, logP, TPSA, etc.), predict drug-likeness (Lipinski's rule), and analyze structural features. The integration abstracts RDKit's API behind a tool interface, allowing the LLM to request analyses by name rather than writing code.

ChemCrow uses a RetryAgentExecutor (from LangChain) that wraps the standard agent executor with retry logic for handling transient failures. When a tool execution fails or the agent reaches an invalid state, the executor retries the operation up to a configurable limit before giving up. This improves robustness in production environments where external services (APIs, databases) may be temporarily unavailable.

ChemCrow uses domain-specific prompts and few-shot examples (embedded in the ChatZeroShotAgent) to guide the LLM toward chemistry-appropriate reasoning. The prompts instruct the LLM to think step-by-step about chemistry problems, consider safety implications, and use available tools appropriately. Few-shot examples demonstrate how to format tool inputs (SMILES, reaction descriptions) and interpret tool outputs, improving the LLM's ability to work with chemistry-specific data formats.

ChemCrow integrates with RXN4Chem (IBM's reaction prediction API) or self-hosted Docker-based reaction engines to predict reaction outcomes and plan synthetic routes. The agent can submit reactant SMILES to the reaction tool, receive predicted products, and iteratively refine synthesis plans. Configuration allows switching between cloud API (RXN4Chem) and local Docker containers via the local_reaction_processing flag, enabling offline operation for sensitive workflows.

ChemCrow includes safety tools that evaluate chemical hazard information, toxicity data, and regulatory compliance for compounds. These tools query safety databases and integrate with the agent to flag dangerous compounds or provide safety recommendations. The safety assessment is integrated into the tool selection logic, allowing the LLM to proactively check safety before recommending synthesis routes or reactions.

ChemCrow integrates paper-qa and PubChem APIs to enable semantic search over chemistry literature and chemical databases. The search tools allow the agent to retrieve relevant papers, chemical data, and synthesis information based on natural language queries. Results are fed back to the LLM for synthesis and summarization, enabling the agent to ground its answers in published research.

ChemCrow provides converter tools that transform between different molecular representation formats (SMILES, IUPAC names, InChI, molecular formulas, etc.). These tools normalize chemical inputs, enabling the agent to work with diverse input formats and convert outputs to user-preferred representations. The converters use RDKit and chemical name resolution libraries to handle ambiguous or non-standard inputs.

ChemCrow allows configuration of separate LLM models for main reasoning (default GPT-4) and tool-specific operations (default GPT-3.5-turbo), with independent temperature settings and max iteration limits. The _make_llm function handles model initialization, supporting both chat and completion models. This dual-model approach optimizes cost (cheaper model for tools) while maintaining reasoning quality (expensive model for planning), with fine-grained control over LLM behavior via temperature (default 0.1 for deterministic chemistry answers).

ChemCrow supports streaming mode (streaming=True) and verbose output (verbose=True) to provide real-time visibility into agent decision-making and tool execution. Streaming returns intermediate results as they become available, while verbose mode logs each tool invocation, LLM reasoning step, and result. This enables debugging, monitoring, and understanding how the agent arrived at its answer through a transparent execution trace.

ChemCrow includes a rephrase chain that post-processes raw tool outputs into coherent, user-friendly answers. After the agent completes tool execution, the rephrase chain reformulates the results using the LLM to improve clarity, remove technical artifacts, and ensure answers are scientifically accurate. This two-stage approach (reasoning + reformulation) decouples tool output quality from answer presentation quality.

ChemCrow implements a modular tool architecture where tools are LangChain BaseTool subclasses organized into categories (RDKit, Reaction, Search, Safety, Converter). Tools are dynamically loaded based on available API keys, allowing graceful degradation when optional services (RXN4Chem, PubChem) are unavailable. The tools.py module provides a factory function that instantiates only available tools, reducing dependencies and enabling flexible deployment across different environments.