Core Engine & Pipeline IMPLEMENTED

The MRP-VM core engine (MRPEngine) orchestrates a 12-phase pipeline for every chat turn. It receives a session-scoped NL request from the server, coordinates normalization, decomposition, retrieval, plugin invocation, and answer synthesis, then commits the turn to session state.

Pipeline Diagram

Phase Details

#	Phase	What Happens	Module	LLM?	Status
1	Receive Request	Server passes { sessionId?, model?, processingMode?, retrievalProfile?, messages[] }	DS013 Server	No	IMPLEMENTED
2	Prepare Turn	Resolve/create session, check expiry, build history, extract current message	DS019 ConversationHandler	No	IMPLEMENTED
3	Resolve Helpers	Built-in manifests resolve explicit seed generators, context normalizers, and response renderers from the SDK helper maps.	DS022 SDK helpers	No	IMPLEMENTED
4	NL → Intent CNL	Normalize current message + history + system prompt into Intent CNL	DS006 Normalizer	Yes	IMPLEMENTED
5	Validate Intent CNL	Structural validation and enum verification of produced Intent CNL	DS007 Validator	No	IMPLEMENTED
6	NL → Session Context CNL	Extract facts, preferences, constraints from current message into Context CNL	DS006 Normalizer	Yes	IMPLEMENTED
7	Validate Session Context	Structural validation and parsing of session context units	DS007 Validator/Parser	No	IMPLEMENTED
8	Decompose Intents	Extract decomposedIntents[] with context profiles from validated Intent CNL	DS011 IntentDecomposer	No	IMPLEMENTED
9	Retrieve Context	Per intent group: derive context profile, run retrieval strategies (BM25, HDC/VSA, etc.)	DS012 Retrieval	No	IMPLEMENTED
10	Invoke Plugins	If needed, dispatch one deterministic plugin per intent group keyed by intentRef	DS003 PluginManager	No	IMPLEMENTED
11	Synthesize Answer	Combine resolved intents + plugin outputs into structured Markdown response	DS017 AnswerSynthesizer	Yes	IMPLEMENTED
12	Commit Turn	Persist user message, assistant response, context units, preferences to session	DS019 ConversationHandler	No	IMPLEMENTED

LLM Budget

Parameter	Value	Notes
Primary LLM stages per request	3	Intent normalization, session-context extraction, synthesis
Max LLM attempts per request	5	Includes corrective retries for validation failures (configurable)
Max correction retries per stage	1	Initial attempt + 1 retry on validation failure
Total timeout	60s	Configurable via requestTimeoutMs
Budget exceeded error	ENGINE_BUDGET_EXCEEDED	No partial return on exhaustion

Failure Handling

Failure	Error Code	Behavior
Intent normalization LLM fails	NORMALIZER_FAILED	Request fails
Intent CNL invalid after retry	NORMALIZER_VALIDATION_FAILED	Request fails
Session context LLM fails	SESSION_CONTEXT_FAILED	Request fails
Session context invalid after retry	SESSION_CONTEXT_VALIDATION_FAILED	Request fails
Zero intent groups	DECOMPOSER_EMPTY_RESULT	Request fails
No evidence for a group	—	Group marked no-context, valid result
Plugin failure/timeout	—	Group marked plugin-error, no fallback
Synthesis fails	SYNTHESIS_FAILED	Request fails, no raw-context fallback

Related Specs

• DS002 — MRP-VM Core Engine
• DS006 — NL Normalizer
• DS007 — CNL Validator & Parser
• DS011 — Intent Decomposition
• DS012 — Retrieval & Context Matching
• DS017 — Answer Synthesis