Skip to main content

Crate cortex_llm

Crate cortex_llm 

Source
Expand description

LLM adapters used by Cortex.

This crate must not write memory rows — it only proposes candidates that the reflection / memory layers ingest (BUILD_SPEC §8). The shared shape lives in adapter; the deterministic CI / fixture adapter lives in replay. Hosted backends (Claude, Ollama) plug into the same adapter::LlmAdapter trait.

Re-exports§

pub use adapter::blake3_hex;
pub use adapter::LlmAdapter;
pub use adapter::LlmError;
pub use adapter::LlmMessage;
pub use adapter::LlmRequest;
pub use adapter::LlmResponse;
pub use adapter::LlmRole;
pub use adapter::TokenUsage;
pub use claude_http::ClaudeHttpAdapter;
pub use claude_http::CLAUDE_ADAPTER_API_KEY_MISSING_INVARIANT;
pub use claude_http::CLAUDE_ADAPTER_ENDPOINT_REJECTED_INVARIANT;
pub use claude_http::CLAUDE_ADAPTER_MODEL_NOT_ALLOWED_INVARIANT;
pub use claude_summary::ClaudeSummaryBackend;
pub use ollama_summary::canonical_prompt_template_blake3;
pub use ollama_summary::OllamaSummaryBackend;
pub use ollama::validate_config as validate_ollama_config;
pub use ollama::validate_endpoint_url as validate_ollama_endpoint_url;
pub use ollama::validate_model_ref as validate_ollama_model_ref;
pub use ollama::OllamaConfig;
pub use ollama_http::OllamaHttpAdapter;
pub use openai_compat::OpenAiCompatAdapter;
pub use replay::FixtureFile;
pub use replay::FixtureIndex;
pub use replay::FixtureMatch;
pub use replay::FixtureResponse;
pub use replay::IndexEntry;
pub use replay::ReplayAdapter;
pub use sensitivity::check_remote_prompt_sensitivity;
pub use sensitivity::MaxSensitivity;
pub use sensitivity::SensitivityGateResult;
pub use summary::NoopSummaryBackend;
pub use summary::ReplaySummaryBackend;
pub use summary::ReplaySummaryFixture;
pub use summary::ReplaySummaryFixtureEntry;
pub use summary::SourceClaim;
pub use summary::SummaryBackend;
pub use summary::SummaryError;
pub use summary::SummaryRequest;
pub use summary::SummaryResponse;

Modules§

adapter
LlmAdapter trait and the request / response / error types it exchanges.
claude_http
HTTP adapter that posts to the Anthropic Messages API.
claude_summary
ClaudeSummaryBackend — Anthropic-backed implementation of SummaryBackend for the Phase 4.D decay LLM-summary path.
ollama
Local validation helpers for Ollama configuration.
ollama_http
HTTP adapter that posts to a local Ollama /api/chat endpoint.
ollama_summary
OllamaSummaryBackend — Ollama-backed implementation of SummaryBackend for the Phase 4.D decay LLM-summary path.
openai_compat
HTTP adapter that posts to any OpenAI-compatible /v1/chat/completions endpoint.
replay
Deterministic LLM adapter that replays canned responses from on-disk fixtures.
sensitivity
Remote-prompt data-classification sensitivity gate.
summary
Pluggable LLM summary backend used by the Phase 4.D decay path.