docs.rs failed to build rullama-seal-0.12.0
Please check the build logs for more information.
See Builds for ideas on how to fix a failed build, or Metadata for how to configure docs.rs builds.
If you believe this is docs.rs' fault, open an issue.

SEAL — Self-Evolving Agentic Learning

This crate (rullama-seal) implements the SEAL framework for enhancing conversational question answering and agent decision-making. It provides coreference resolution, structured query extraction, self-evolving pattern learning, and post-execution reflection — enabling agents to understand implicit references, build reusable knowledge, and correct their own mistakes without retraining.

Inspired by: SEAL: Self-Evolving Agentic Learning for Conversational Question Answering over Knowledge Graphs (Wang et al., arXiv:2512.04868, December 2024)

Crate boundary

SEAL spent part of the 0.10 cycle folded into rullama-agent behind a seal feature flag; in 0.11 it moved back out into the standalone rullama-seal crate. The dependencies it needs — ResponseConfidence (now in rullama-core), ToolOutcome / ToolErrorCategory (in rullama-tool-runtime), RelationshipGraph (in rullama-knowledge) — are all addressable from a leaf crate, so the standalone shape is back to being the right one. Optional integrations use the knowledge, feedback, and mdap features.

Feature Flags

Feature	Description
`seal`	Core SEAL pipeline (coreference, query extraction, learning, reflection)
`seal-mdap`	MDAP metric recording via `mdap` feature
`seal-knowledge`	BKS/PKS knowledge system integration via `rullama-knowledge`
`seal-feedback`	Audit feedback bridge via `rullama-permission`

# Core SEAL
rullama-seal = "0.12"

# With knowledge integration
rullama-seal = { version = "0.12", features = ["knowledge"] }

# Via the rullama facade
rullama = { version = "0.12", features = ["seal"] }

Architecture

User Query
    │
    ▼
┌─── Coreference Resolution ─────────────────────────────────────┐
│  detect_references() → resolve() → rewrite_with_resolutions()  │
│  "What uses it?" → "What uses [main.rs]?"                      │
└────────────────────────────────┬────────────────────────────────┘
                                 │
                                 ▼
┌─── Query Core Extraction ──────────────────────────────────────┐
│  classify() → build_expression() → QueryCore                   │
│  S-expression: (JOIN DependsOn ?dep "main.rs")                 │
└────────────────────────────────┬────────────────────────────────┘
                                 │
                                 ▼
┌─── Learning Coordinator ───────────────────────────────────────┐
│  Local Memory (per-session)  │  Global Memory (cross-session)  │
│  process_query() → match pattern or create new                 │
│  record_outcome() → update reliability scores                  │
└────────────────────────────────┬────────────────────────────────┘
                                 │
                                 ▼
┌─── Reflection Module ──────────────────────────────────────────┐
│  analyze() → detect issues → suggest fixes → attempt_correction│
│  Errors: EmptyResult, Overflow, EntityNotFound, RelationMismatch│
└────────────────────────────────────────────────────────────────┘

Quick Start

use rullama_seal::{SealProcessor, SealConfig, DialogState};
use rullama_core::graph::{EntityStoreT, RelationshipGraphT};

let mut processor = SealProcessor::with_defaults();
processor.init_conversation("session-001");

let result = processor.process(
    "What uses it?",
    &dialog_state,
    &entity_store,
    Some(&graph),
)?;

println!("Resolved: {}", result.resolved_query);

Components