deepadata-langchain-adapter

Wraps LangChain memory implementations with an Emotional Data Model (EDM) layer that annotates conversation contexts with governed emotional metadata (valence, arousal, sentiment dimensions). The adapter intercepts memory read/write operations, applies schema-based emotional classification to messages, and stores enriched tuples of (content, emotional_vector, timestamp) without modifying underlying LangChain memory backends, enabling downstream reasoning over emotional state trajectories.

Provides a unified interface that works across multiple LangChain memory types (BufferMemory, ConversationSummaryMemory, EntityMemory, etc.) by implementing a generic wrapper that intercepts the memory interface contract. The adapter detects the underlying memory type, applies EDM enrichment at the abstraction layer, and delegates actual storage to the original memory implementation, enabling emotional tracking regardless of memory strategy chosen.

Converts conversation messages and context into multi-dimensional emotional state vectors using the EDM schema, producing numeric representations of emotional dimensions (e.g., valence, arousal, sentiment polarity, emotional intensity). The adapter applies a configurable classifier (LLM-based, rule-based, or hybrid) to extract emotional features from text, normalizes them to a standard vector space, and stores the vectors alongside memory entries for downstream reasoning or similarity comparisons.

Enhances LangChain memory retrieval by filtering or ranking memory entries based on emotional context similarity. When retrieving relevant past messages, the adapter compares the current emotional state vector against stored emotional vectors, optionally re-ranking results by emotional relevance or filtering to emotionally similar contexts. This enables the agent to retrieve not just semantically relevant memories but emotionally resonant ones, improving response coherence in emotionally-charged conversations.

Provides configuration interfaces to customize the Emotional Data Model schema used for enrichment, allowing developers to define which emotional dimensions are tracked, how they are classified, and how they are normalized. The adapter exposes schema definition objects that map emotional concepts to classifier prompts or rules, enabling teams to adapt the EDM to domain-specific emotional vocabularies (e.g., customer support vs. mental health vs. gaming contexts) without forking the adapter code.

Provides hooks and utilities to persist emotional metadata alongside memory entries when using LangChain's save/load mechanisms. The adapter intercepts serialization events, includes emotional vectors and EDM annotations in the persisted data, and reconstructs emotional state when memory is loaded from storage. This enables emotional context to survive across sessions and deployments, supporting long-term emotional trajectory analysis and continuity in multi-session conversations.

Monitors emotional state vectors across consecutive messages to detect significant emotional transitions or anomalies (e.g., sudden shift from calm to frustrated, sustained high arousal). The adapter compares emotional vectors using configurable distance metrics and thresholds, emits events when state changes exceed thresholds, and optionally triggers callbacks or logging. This enables agents to react to emotional shifts in real-time, such as escalating to a human agent when user frustration spikes.

Analyzes emotional state across multiple conversation turns to compute trajectory metrics such as emotional stability, trend direction (improving/deteriorating), and cumulative emotional load. The adapter aggregates emotional vectors over conversation windows, applies statistical measures (variance, slope, moving averages), and produces summary metrics describing the overall emotional arc of a conversation. This enables understanding of long-term emotional patterns rather than isolated snapshots.