dohertj2/chat
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

merge: T43 multi-entity prompt assembly

Browse Source
This commit is contained in:
Joseph Doherty
2026-04-26 16:08:54 -04:00
parent a90647dddb fcb111310a
commit 9b601650fb
2 changed files with 366 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files
chat/services/prompt.py
+125 -35
View File
@@ -37,6 +37,7 @@ import tiktoken
from chat.llm.client import Message
from chat.state.edges import get_edge, list_edges_for
from chat.state.entities import get_bot, get_you
from chat.state.group_node import get_group_node
from chat.state.memory import search_memories
from chat.state.world import (
active_scene,
@@ -206,6 +207,26 @@ def _build_previous_scene_block(pov_summary: str | None) -> str | None:
return "PREVIOUS SCENE SUMMARY:\n" + pov_summary
def _build_group_node_block(group_node: dict | None) -> str | None:
"""Render the group-node summary + dynamic as a SHOULD-tier block.
Used only in 3-entity scenes (you + host + guest). Returns None when
the row is missing or both summary and dynamic are empty.
"""
if not group_node:
return None
summary = (group_node.get("summary") or "").strip()
dynamic = (group_node.get("dynamic") or "").strip()
if not summary and not dynamic:
return None
lines = ["Group dynamic:"]
if summary:
lines.append(f"- Summary: {summary}")
if dynamic:
lines.append(f"- Dynamic: {dynamic}")
return "\n".join(lines)
def _closing_instruction(speaker_name: str, addressee_name: str) -> str:
return (
f"Continue the scene as {speaker_name}, in their voice, responding "
@@ -287,6 +308,7 @@ def assemble_narrative_prompt(
budget_soft: int = 6000,
budget_hard: int = 8000,
encoding_name: str = "cl100k_base",
guest_id: str | None = None,
) -> list[Message]:
"""Assemble the narrative prompt for ``speaker_bot_id`` to respond.
@@ -313,6 +335,15 @@ def assemble_narrative_prompt(
if chat is None:
raise ValueError(f"chat_id {chat_id!r} not found")
# Auto-detect guest from chat state when caller didn't pass one.
# Phase 1 chats have ``guest_bot_id is None``; the auto-detect is a
# no-op there and the function behaves exactly as before.
if guest_id is None:
guest_id = chat.get("guest_bot_id")
# A speaker addressing themself as guest doesn't add a third party.
if guest_id is not None and guest_id == speaker_bot_id:
guest_id = None
you = get_you(conn)
addressee_id, addressee_name = _resolve_addressee(conn, addressee, you)
@@ -325,9 +356,10 @@ def assemble_narrative_prompt(
addressee_name,
)
# Activity for present entities. Phase 1: you + speaker bot. (When a
# guest is added in Phase 1+, callers that know about it can pass
# extra activities via a future hook; for now we keep it strict.)
# Activity for present entities. Core (MUST): you + speaker bot.
# Phase 2 (SHOULD-tier): when a third party (guest) is present in
# the chat, append their activity in a separate block so it can be
# trimmed independently under tight budget.
activities: list[dict] = []
you_act = get_activity(conn, "you")
if you_act is not None:
@@ -341,6 +373,34 @@ def assemble_narrative_prompt(
activities.append(bot_act)
activity_block = _build_activity_block(activities)
# SHOULD-tier guest activity extension (Phase 2 / Task 43).
guest_activity_block: str | None = None
if guest_id is not None:
guest_act = get_activity(conn, guest_id)
if guest_act is not None:
guest_act = dict(guest_act)
guest_bot = get_bot(conn, guest_id)
guest_act["_display_name"] = (
guest_bot["name"] if guest_bot else guest_id
)
guest_activity_block = _build_activity_block([guest_act])
# SHOULD-tier group-node block (Phase 2 / Task 43): rendered only
# when the group_node row is present AND it covers all three of
# you + host + guest (per the Task 43 spec).
group_node_block: str | None = None
if guest_id is not None:
gn = get_group_node(conn, chat_id)
if gn is not None:
members = set(gn.get("members") or [])
host_id = chat.get("host_bot_id")
required = {"you"}
if host_id is not None:
required.add(host_id)
required.add(guest_id)
if required.issubset(members):
group_node_block = _build_group_node_block(gn)
container = None
if chat.get("active_scene_id"):
scene = get_scene(conn, chat["active_scene_id"])
@@ -421,6 +481,8 @@ def assemble_narrative_prompt(
include_previous_scene: bool,
include_memories_top_k: int,
dialogue_keep: int,
include_guest_activity: bool = True,
include_group_node: bool = True,
) -> tuple[str, int, list[dict]]:
# dialogue: keep the last `dialogue_keep` turns verbatim; older
# turns become an "earlier:" placeholder line.
@@ -447,6 +509,8 @@ def assemble_narrative_prompt(
other_edges_block if include_other_edges else None,
scene_block,
activity_block,
guest_activity_block if include_guest_activity else None,
group_node_block if include_group_node else None,
prev_block,
memories_block,
dialogue_block,
@@ -463,12 +527,25 @@ def assemble_narrative_prompt(
nice_memories_k = min(4, len(memory_summaries))
include_prev = previous_scene_summary is not None
include_other = other_edges_block is not None
include_guest_activity = guest_activity_block is not None
include_group_node = group_node_block is not None
body, total, _ = assemble(
include_other_edges=include_other,
include_previous_scene=include_prev,
include_memories_top_k=nice_memories_k,
dialogue_keep=nice_dialogue_keep,
def _build(*, prev: bool, mem_k: int, dlg: int, other: bool,
guest_act: bool, group: bool) -> tuple[str, int]:
body, total, _ = assemble(
include_other_edges=other,
include_previous_scene=prev,
include_memories_top_k=mem_k,
dialogue_keep=dlg,
include_guest_activity=guest_act,
include_group_node=group,
)
return body, total
body, total = _build(
prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
other=include_other, guest_act=include_guest_activity,
group=include_group_node,
)
# If under soft, we're done.
@@ -478,34 +555,31 @@ def assemble_narrative_prompt(
# Drop NICE in order: previous scene → memories beyond top-2 →
# older dialogue turns (collapse to 4).
if include_prev:
body, total, _ = assemble(
include_other_edges=include_other,
include_previous_scene=False,
include_memories_top_k=nice_memories_k,
dialogue_keep=nice_dialogue_keep,
)
include_prev = False
body, total = _build(
prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
other=include_other, guest_act=include_guest_activity,
group=include_group_node,
)
if total <= budget_soft:
return _emit(body, user_turn_prose)
if nice_memories_k > 2:
nice_memories_k = 2
body, total, _ = assemble(
include_other_edges=include_other,
include_previous_scene=False,
include_memories_top_k=nice_memories_k,
dialogue_keep=nice_dialogue_keep,
body, total = _build(
prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
other=include_other, guest_act=include_guest_activity,
group=include_group_node,
)
if total <= budget_soft:
return _emit(body, user_turn_prose)
if nice_dialogue_keep > baseline_keep:
nice_dialogue_keep = baseline_keep
body, total, _ = assemble(
include_other_edges=include_other,
include_previous_scene=False,
include_memories_top_k=nice_memories_k,
dialogue_keep=nice_dialogue_keep,
body, total = _build(
prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
other=include_other, guest_act=include_guest_activity,
group=include_group_node,
)
if total <= budget_soft:
return _emit(body, user_turn_prose)
@@ -513,21 +587,37 @@ def assemble_narrative_prompt(
# Drop more NICE until we're under hard: memories all the way to 0.
while nice_memories_k > 0 and total > budget_hard:
nice_memories_k = max(0, nice_memories_k - 1)
body, total, _ = assemble(
include_other_edges=include_other,
include_previous_scene=False,
include_memories_top_k=nice_memories_k,
dialogue_keep=nice_dialogue_keep,
body, total = _build(
prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
other=include_other, guest_act=include_guest_activity,
group=include_group_node,
)
# Drop SHOULD-tier blocks in order: guest activity → group node →
# other edges. (Guest activity goes first per Task 43 spec — it's
# the most expendable additive context.)
if include_guest_activity and total > budget_hard:
include_guest_activity = False
body, total = _build(
prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
other=include_other, guest_act=include_guest_activity,
group=include_group_node,
)
if include_group_node and total > budget_hard:
include_group_node = False
body, total = _build(
prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
other=include_other, guest_act=include_guest_activity,
group=include_group_node,
)
# Drop SHOULD: other edges.
if include_other and total > budget_hard:
include_other = False
body, total, _ = assemble(
include_other_edges=False,
include_previous_scene=False,
include_memories_top_k=nice_memories_k,
dialogue_keep=nice_dialogue_keep,
body, total = _build(
prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
other=include_other, guest_act=include_guest_activity,
group=include_group_node,
)
if total > budget_hard:
tests/test_prompt.py
+241
View File
@@ -253,3 +253,244 @@ def test_must_exceeds_budget_hard_raises_value_error(tmp_path):
budget_soft=5,
budget_hard=10,
)
# ---------------------------------------------------------------------------
# Task 43: multi-entity prompt assembly (guest_id support)
# ---------------------------------------------------------------------------
def _seed_with_guest(conn) -> None:
"""Seed a 3-entity scene: you (Sam) + host (Aria, bot_a) + guest (Iris, bot_b).
Group node row is initialized with summary + dynamic, edges in all
relevant directions are seeded, and activities are recorded for all
three entities.
"""
append_event(conn, kind="bot_authored", payload={
"id": "bot_a",
"name": "Aria",
"persona": "reserved coworker who notices things",
"voice_samples": ["I — sorry, I didn't mean to.", "Right. Of course."],
"traits": ["introverted", "observant"],
"backstory": "An archivist who joined the firm last spring.",
"initial_relationship_to_you": "coworker; mild crush; never voiced",
"kickoff_prose": "you stay late at the office",
})
append_event(conn, kind="bot_authored", payload={
"id": "bot_b",
"name": "Iris",
"persona": "wry transplant from the Boston office",
"voice_samples": ["Oh, please.", "Don't make me say it twice."],
"traits": ["sardonic", "loyal"],
"backstory": "Met Aria at a conference two years back.",
"initial_relationship_to_you": "stranger; curious",
"kickoff_prose": "",
})
append_event(conn, kind="you_authored", payload={
"name": "Sam",
"pronouns": "they/them",
"persona": "tired analyst",
})
append_event(conn, kind="chat_created", payload={
"id": "chat_bot_a",
"host_bot_id": "bot_a",
"guest_bot_id": "bot_b",
"initial_time": "2026-04-26T20:00:00+00:00",
"narrative_anchor": "Day 1 evening",
"weather": "clear",
})
append_event(conn, kind="container_created", payload={
"chat_id": "chat_bot_a",
"name": "office bullpen",
"type": "workplace",
"properties": {"public": False, "moving": False, "audible_range": "room"},
})
# Edges: host -> you, guest -> you, host -> guest, guest -> host.
append_event(conn, kind="edge_update", payload={
"source_id": "bot_a",
"target_id": "you",
"affinity_delta": 12,
"trust_delta": 5,
"knowledge_facts": ["they work on the same floor"],
})
append_event(conn, kind="edge_update", payload={
"source_id": "bot_a",
"target_id": "bot_b",
"affinity_delta": 20,
"trust_delta": 15,
"knowledge_facts": ["studied physics together"],
})
append_event(conn, kind="edge_update", payload={
"source_id": "bot_b",
"target_id": "you",
"affinity_delta": 4,
"trust_delta": 0,
"knowledge_facts": ["Aria's coworker"],
})
append_event(conn, kind="edge_update", payload={
"source_id": "bot_b",
"target_id": "bot_a",
"affinity_delta": 18,
"trust_delta": 12,
"knowledge_facts": ["former roommate"],
})
# Activity for all three entities — note distinct verbs so we can
# check whose activity got dropped under tight budget.
append_event(conn, kind="activity_change", payload={
"entity_id": "you",
"container_id": 1,
"posture": "sitting at your desk",
"action": {"verb": "finishing emails"},
"attention": "the screen",
"holding": ["coffee mug"],
})
append_event(conn, kind="activity_change", payload={
"entity_id": "bot_a",
"container_id": 1,
"posture": "sitting at her desk",
"action": {"verb": "pretending to work"},
"attention": "you, in glances",
})
append_event(conn, kind="activity_change", payload={
"entity_id": "bot_b",
"container_id": 1,
"posture": "leaning against the doorframe",
"action": {"verb": "smirking-distinctively"},
"attention": "Aria",
})
append_event(conn, kind="scene_opened", payload={
"chat_id": "chat_bot_a",
"container_id": 1,
"started_at": "2026-04-26T20:00:00+00:00",
"participants": ["you", "bot_a", "bot_b"],
})
append_event(conn, kind="group_node_initialized", payload={
"chat_id": "chat_bot_a",
"members": ["you", "bot_a", "bot_b"],
"summary": "Three coworkers catching up after hours UNIQUE-GROUP-SUMMARY.",
"dynamic": "warm-but-prickly UNIQUE-GROUP-DYNAMIC",
})
project(conn)
def test_assemble_with_no_guest_matches_phase1(tmp_path):
"""Regression: 2-entity scenario without guest_id behaves exactly as Phase 1."""
db = tmp_path / "t.db"
apply_migrations(db)
with open_db(db) as conn:
_seed_basic(conn)
msgs = assemble_narrative_prompt(
conn,
chat_id="chat_bot_a",
speaker_bot_id="bot_a",
recent_dialogue=[],
retrieved_memory_summaries=[],
)
body = msgs[0].content
# Phase 1 must blocks present.
assert "Aria" in body
assert "PERSONA" in body
assert "Sam" in body
assert "ACTIVITIES" in body
assert "62/100" in body # speaker → addressee edge intact
# No guest content leaks in.
assert "Group dynamic" not in body
assert "Iris" not in body
def test_assemble_with_guest_includes_group_node_summary(tmp_path):
"""When guest is present (auto-detected via chat.guest_bot_id) and a
group_node row exists, its summary + dynamic are rendered."""
db = tmp_path / "t.db"
apply_migrations(db)
with open_db(db) as conn:
_seed_with_guest(conn)
msgs = assemble_narrative_prompt(
conn,
chat_id="chat_bot_a",
speaker_bot_id="bot_a",
recent_dialogue=[],
retrieved_memory_summaries=[],
)
body = msgs[0].content
assert "Group dynamic" in body
assert "UNIQUE-GROUP-SUMMARY" in body
assert "UNIQUE-GROUP-DYNAMIC" in body
# Guest activity also present (SHOULD-tier, fits at default budget).
assert "smirking-distinctively" in body
# Speaker's other edges include the host -> guest direction.
assert "Iris" in body
def test_assemble_when_speaker_is_guest_orients_edges_correctly(tmp_path):
"""When the guest is the speaker, identity is the guest, the
addressee edge is guest → you, and other edges include guest → host."""
db = tmp_path / "t.db"
apply_migrations(db)
with open_db(db) as conn:
_seed_with_guest(conn)
msgs = assemble_narrative_prompt(
conn,
chat_id="chat_bot_a",
speaker_bot_id="bot_b", # guest as speaker
recent_dialogue=[],
retrieved_memory_summaries=[],
)
body = msgs[0].content
# Speaker identity is the guest's persona.
assert "You are Iris." in body
assert "wry transplant from the Boston office" in body
# Edge to addressee is guest → you (Sam) with the seeded values
# (default 50 + 4 affinity = 54).
assert "YOUR EDGE TO Sam" in body
assert "54/100" in body
# Other edges include guest → host (Aria) with seeded value
# (default 50 + 18 = 68).
assert "OTHER EDGES" in body
assert "Aria" in body
assert "68/100" in body
def test_assemble_with_tight_budget_drops_guest_activity_first(tmp_path):
"""Under tight budget MUST blocks survive but SHOULD-tier guest
activity is dropped first."""
db = tmp_path / "t.db"
apply_migrations(db)
with open_db(db) as conn:
_seed_with_guest(conn)
# Short dialogue so MUST core (speaker identity + edge + last 4
# turns + closing) sits comfortably under the hard budget while
# SHOULD-tier additions (guest activity, group node, other edges)
# would push over.
dialogue = [
{"speaker": "you", "text": "line-16 hi there"},
{"speaker": "bot_a", "text": "line-17 hey"},
{"speaker": "you", "text": "line-18 quiet night"},
{"speaker": "bot_a", "text": "line-19 indeed"},
]
msgs = assemble_narrative_prompt(
conn,
chat_id="chat_bot_a",
speaker_bot_id="bot_a",
recent_dialogue=dialogue,
retrieved_memory_summaries=[],
# MUST core ~310 tokens; SHOULD additions (guest activity +
# group node + other edges) push it well over 380. budget_hard
# is set just above MUST core so SHOULD-tier blocks must be
# trimmed away.
budget_soft=250,
budget_hard=340,
)
body = msgs[0].content
# MUST: speaker identity, edge to addressee, last 4 dialogue turns.
assert "Aria" in body
assert "YOUR EDGE TO Sam" in body
for i in range(16, 20):
assert f"line-{i:02d}" in body
# Guest activity (SHOULD-tier) must be dropped under tight budget.
assert "smirking-distinctively" not in body
# Token budget honoured.
import tiktoken
enc = tiktoken.get_encoding("cl100k_base")
assert len(enc.encode(body)) <= 340