Foresight.Prompts.Retain.V1 (Foresight v0.1.0)

Copy Markdown View Source

Retain-time atomic fact extraction prompt.

Faithful port of Hindsight's engine/retain/fact_extraction.py (_BASE_FACT_EXTRACTION_PROMPT + _CONCISE_GUIDELINES + _CONCISE_EXAMPLES) and the ExtractedFact schema. The LLM reads a single input (a chat session, document, or paragraph) plus its Event Date and emits a LIST of atomic facts — one input to N discrete memory units — which is the mechanism by which cross-session counts, totals, and temporal ordering become answerable at reflect time. Without it, retain stores whole-session blobs and reflect has nothing structured to aggregate over.

build_fact_text/1 mirrors Hindsight's LIVE inline assembler in _extract_facts_from_chunk (what + optional When: when + Involving: who + why, joined by |) — NOT the dead ExtractedFact.build_fact_text Pydantic method, which omits when and is never called by the real pipeline.

Summary

Functions

build_fact_text(fact)

Combine an extracted fact's dimensions into one stored fact string.

default_mode()

The default extraction mode when a mode is enabled without a specific value.

entity_types()

Entity types offered to the typed-extraction schema/prompt (B3).

messages(input, event_date, opts \\ [])

Build the system + user messages for one extraction call.

modes()

Known extraction modes. nil/"none" disables extraction (passthrough).

schema(opts \\ [])

JSON schema for the facts list response contract (mirrors FactExtractionResponse).

Functions

build_fact_text(fact)

@spec build_fact_text(map()) :: String.t()

Combine an extracted fact's dimensions into one stored fact string.

Faithful to Hindsight's LIVE assembler in _extract_facts_from_chunk (fact_extraction.py ~L1107-1118): what, then When: <when>, then Involving: <who>, then why — each appended only when present and not "N/A", joined by |.

NOTE: this deliberately does NOT mirror the Pydantic ExtractedFact.build_fact_text method (py ~L166-181), which omits when. That method is DEAD CODE in Hindsight — the real retain pipeline never calls it and assembles combined_text inline WITH the When: clause. Porting the method dropped the temporal marker from every stored fact, starving semantic recall and reflect of dates on temporal/multi-session questions. Keep When: here.

default_mode()

@spec default_mode() :: String.t()

The default extraction mode when a mode is enabled without a specific value.

entity_types()

@spec entity_types() :: [String.t()]

Entity types offered to the typed-extraction schema/prompt (B3).

messages(input, event_date, opts \\ [])

@spec messages(String.t(), String.t() | nil, keyword()) :: [map()]

Build the system + user messages for one extraction call.

input is the raw content to atomize. event_date is the ISO/string date the input was mentioned (used to resolve relative temporal expressions). mission steers what is worth extracting (nil → none). mode selects the guideline block; custom_instructions supplies the guidelines for "custom" mode.

modes()

@spec modes() :: [String.t()]

Known extraction modes. nil/"none" disables extraction (passthrough).

schema(opts \\ [])

@spec schema(keyword()) :: map()

JSON schema for the facts list response contract (mirrors FactExtractionResponse).

Opts:

  • :typed_entities — when true, entities items are {name, type} objects with a type enum (B3 exceed) instead of bare strings. Default false (parity).