Cohere: Command R7B (12-2024)

Implements RAG by accepting external document contexts and ranking them based on relevance to the query before generation, using a learned ranking mechanism that weights document importance during token generation. The model integrates retrieved context directly into the prompt context window, allowing it to synthesize answers grounded in provided documents while maintaining coherence across multiple sources.

Supports structured tool invocation through a schema-based function registry where tools are defined as JSON schemas with parameters, descriptions, and return types. The model generates tool calls as structured JSON that can be routed to external APIs or local functions, with built-in support for multi-turn tool use where results are fed back into the conversation context for further reasoning.

Follows complex, multi-part instructions with high fidelity, respecting constraints on output format, length, style, and content restrictions. The model is trained to parse and execute detailed prompts, maintaining compliance across multiple simultaneous constraints and handling edge cases gracefully.

Maintains conversation history across multiple turns with full context preservation, allowing the model to reference previous exchanges, build on prior reasoning, and correct itself based on feedback. The model uses a sliding context window that prioritizes recent messages while optionally summarizing or truncating older turns to stay within token limits.

Supports explicit reasoning chains where the model breaks down complex problems into intermediate steps, showing work before arriving at conclusions. This is implemented through prompt-level instruction for step-by-step reasoning, combined with the model's training on reasoning tasks, enabling it to handle multi-hop logical inference, mathematical problem-solving, and structured decision-making.

Generates coherent, contextually appropriate text across multiple styles and tones through instruction-based control, where prompts can specify desired voice (formal, casual, technical, creative), length constraints, and output format. The model uses instruction-tuning to respect these constraints while maintaining semantic accuracy and coherence.

Extracts structured information (entities, relationships, attributes) from unstructured text by accepting JSON schema definitions and returning parsed data matching those schemas. The model performs entity recognition, relationship extraction, and attribute assignment through instruction-tuned prompting, with support for nested structures and optional fields.

Generates code snippets, complete functions, and multi-file solutions in multiple programming languages through instruction-based prompting. The model understands code context, can refactor existing code, and provides explanations alongside generated code, leveraging its training on diverse codebases and technical documentation.

Condenses long documents or conversations into summaries of varying lengths and detail levels, from single-sentence abstracts to detailed bullet-point summaries. The model uses instruction-based control to balance comprehensiveness with brevity, preserving key information while removing redundancy.

Generates and translates text across multiple languages with support for context-aware localization. The model understands cultural nuances and can adapt content for different linguistic contexts, though translation quality varies by language pair and domain.

Ranks text passages or documents by relevance to a query through semantic understanding, without explicit vector embeddings. The model evaluates semantic similarity by processing both query and candidates in context, producing relevance scores that reflect deeper semantic relationships than keyword matching.