galet-prompt-builder

Provider-neutral orchestration of CoALA memory into bounded prompts.

The package compiles caller-supplied system instructions and current input with procedural, episodic, and semantic memory from galet-memory. It applies independent section budgets, a total hard limit, relevance assessment, digest deduplication, deterministic selection, and detailed metrics.

It deliberately has no concept of Lucy agents, handlers, storage paths, attachments, or application configuration. Lucy can later adapt those values into PromptRequest, PromptBudgets, and PromptLimits.

Basic use

from galet_prompt_builder import (
    PromptBudgets,
    PromptCompiler,
    PromptLimits,
    PromptRequest,
)

compiler = PromptCompiler(
    procedural_memory=procedural_memory,
    episodic_memory=episodic_memory,
    semantic_memory=semantic_memory,
)

compiled = compiler.compile(
    PromptRequest(
        account_name="demo",
        conversation_id="session-id",
        context_name="project",
        system_instructions=["You are a careful assistant."],
        current_input="What did we decide about storage?",
        semantic_namespaces=("documents",),
        semantic_score_threshold=0.30,
        episodic_event_kinds=(
            "user_message",
            "assistant_message",
            "session_digest",
        ),
        include_structured_episodic_events=False,
        episodic_digest_score_threshold=0.40,
    ),
    PromptBudgets(
        total_tokens=8000,
        procedural_tokens=1500,
        episodic_event_tokens=2000,
        episodic_digest_tokens=1000,
        semantic_tokens=2000,
        safety_margin_tokens=500,
    ),
    PromptLimits(
        maximum_total_tokens=16000,
        maximum_procedural_tokens=3000,
        maximum_episodic_event_tokens=4000,
        maximum_episodic_digest_tokens=2000,
        maximum_semantic_tokens=4000,
        maximum_events=6,
        maximum_digests=2,
        maximum_semantic_documents=3,
        maximum_episodic_event_chars=4000,
        maximum_episodic_digest_chars=1800,
        maximum_semantic_item_chars=1800,
    ),
)

provider_messages = compiled.provider_messages
metrics = compiled.metrics

The default relevance assessor is deterministic and makes no network calls. GaletRelevanceAssessor is available when an application explicitly wants an LLM-based assessment; all such calls go through Galet's LLMApi.

Prompt policy is explicit and agent-independent. Applications choose the overall and per-section token budgets, retrieval counts, event kinds, relevance thresholds, and per-item character caps. An agent may supply system instructions, but no agent object or application policy is required.

Prompt comparison CLI

The installed galet-prompt-run command compiles a prompt from an existing Lucy chat and a new request. It resolves the chat's friendly name to its stored session ID and reads episodic history from Lucy's SQLite database.

By default it reads:

/home/junwin/lucy_storage/data/chat2.sqlite

Example:

galet-prompt-run \
  "What should we do next?" \
  --chat-name "Prompt Builder Work"

Use a different storage root or an explicit database when needed:

galet-prompt-run \
  "Compare this prompt" \
  --chat-name "Prompt Builder Work" \
  --storage-root /srv/lucy_storage

galet-prompt-run \
  "Compare this prompt" \
  --chat-name "Prompt Builder Work" \
  --db /tmp/chat2.sqlite \
  --format json

The command is read-only. It fails if the database or friendly chat name does not exist, and it reports duplicate friendly names instead of guessing. The runner includes episodic events and digests by default.

Semantic recall is opt-in. Supply one or more namespaces to query Lucy's embeddings-v2.sqlite store:

galet-prompt-run \
  "What did I write about attention?" \
  --chat-name "Prompt Builder Work" \
  --namespaces vol_6 vol_7 documents

The comparison runner uses deliberately compact defaults:

• 8,000 total tokens with a 500-token safety margin
• 1,000 recent-event tokens across at most 6 events
• 500 digest tokens across at most 2 digests
• 1,000 semantic tokens across at most 3 documents
• semantic score threshold 0.30
• digest score threshold 0.40
• conversational event kinds only; structured tool payloads are excluded

Every policy value can be stated explicitly:

galet-prompt-run \
  "What did I write about attention?" \
  --chat-name "Prompt Builder Work" \
  --namespaces vol_6 vol_7 documents \
  --total-tokens 6000 \
  --safety-margin-tokens 500 \
  --episodic-event-tokens 800 \
  --max-events 6 \
  --episodic-digest-tokens 400 \
  --max-digests 2 \
  --digest-score-threshold 0.45 \
  --semantic-tokens 900 \
  --semantic-top-k 3 \
  --semantic-score-threshold 0.35 \
  --semantic-item-max-chars 1800

With no --namespaces, the namespace list is empty and the runner does not open the embedding database or make an embedding API call. Semantic mode defaults to /home/junwin/lucy_storage/data/embeddings-v2.sqlite; use --embedding-db to override it and install the optional dependency with pip install ".[semantic]". Galet supplies the query embedding, using its normal credential configuration or the directory passed to --credential-path.

Procedural memory remains disabled until its context repository is configured explicitly.

The text report includes a content-safe candidate decision table. It shows each candidate ID, source, raw relevance score, original/final token cost, and whether it was selected or rejected by the relevance threshold or token budget. The JSON report exposes the same data under metrics.candidates. Rejected candidates still count as retrieved and dropped in their section metrics, making threshold tuning visible without copying document or chat content into diagnostics.

Dependency rule

Applications may depend on galet-prompt-builder; the package must never depend on an application.

The package depends on the neutral interfaces from galet-memory. It never opens memory databases itself and prompt compilation never mutates memory.