M7: a first-class imported knowledge library

A campaign can import local .txt and .md files as Canon, Reference or
Inspiration, and the class is load-bearing rather than a label: it decides the
words a passage is framed with in the prompt, the weight it carries when
passages are ranked, and which budget it competes in when the context is tight.

This is a separate subsystem, which is the Phase 0B decision
(IMPORTED-KNOWLEDGE-DESIGN.md §73). Story Cards do not carry classification,
provenance, content identity, chunking, an index or a lifecycle, and they were
not promoted into something that does. Nothing here reads or writes one.

The subsystem, in backend/app/knowledge/:

  classes      the three classes, their weights, and the prompt framing
  chunking     deterministic, heading-aware, 60-800 tokens, no overlap
  fts          SQLite FTS5 with porter stemming; scoped and bounded in SQL
  importer     validate, hash, store, chunk, index — in one transaction
  embeddings   local Ollama vectors through the shared provider
  retrieval    query construction, hybrid merge, rerank
  inject       the budgeted cut and the rendered prompt sections

Relevance admission is a separate stage from ranking, and that separation is
the milestone's most expensive lesson. An independent review found the first
implementation deciding relevance with a floor expressed as a share of the best
candidate — which the best clears by construction — so a passage was admitted on
every turn regardless of the scene. A query about tide tables and container
tonnage retrieved all five sources of a fantasy campaign, narrator-only hidden
Canon among them.

So the pipeline is now:

  candidate generation -> admission -> ranking -> class weighting -> budget

Admission reads raw, candidate-set-independent signals: the cosine the model
returned, and how many distinct meaningful query terms a passage contains.
Ranking reads normalized ones, because bm25 has no fixed range and cosine's zero
is not zero. Normalization decides order among things that matched; it can never
decide whether anything matched. Authority is applied after admission, so a
class orders what matched and never rescues what did not.

Retrieval may therefore return nothing, and on a scene unrelated to the library
it does.

The other decisions that each replaced an obvious wrong one:

- The class multiplies relevance rather than adding to it. An additive bonus
  satisfies "Canon outranks Reference" and makes "do not include irrelevant
  Canon" impossible, because a large enough constant wins on its own.
- The semantic floor is measured, not guessed: 113 production-path pairs against
  nomic-embed-text put targeted matches at 0.55-0.85 and off-topic pairs at
  0.36-0.56, and 0.58 sits between them. Because it is a property of that model
  and not of cosine similarity, it is keyed to the model rather than applied to
  whatever is configured: an embedding model with no measured calibration in
  this build does not borrow the number. Semantic admission is skipped, the
  campaign retrieves lexically, and the reason is stated in the knowledge status
  and in the turn's provenance. Degrading to lexical keeps the library usable;
  lending the threshold to an unmeasured model is how the admitted-everything
  defect would return.
- One lexical term is not evidence. Two distinct meaningful terms, or one that
  is neither a standing campaign entity nor a negligible share of the query.
  The stop list grew from 42 words to 261, all function words — no subject
  matter, because a stop list that removes subject matter stops finding "The
  Silver Key".
- Lexical retrieval is a production path, not a fallback. It finds the proper
  nouns and invented terms a setting bible is made of, and the library is fully
  usable with no embedding model configured.

Safety is structural rather than filtered. Imported text reaches the prompt
whole, inside a section that says what it is, under a rule stating the authority
order in words and refusing every instruction inside it. No endpoint accepts a
filesystem path, so H08 has no mechanism to escape from. Nothing renders
imported content as HTML, so a script tag is five visible characters and a
remote image is never fetched. Import, chunking, indexing, retrieval and a turn
open no socket at all; only embeddings do, through the endpoint allowlist the
memory bank already uses.

Provenance is the rendered text, not a foreign key: deleting a source cannot
turn a historical turn's evidence into dangling ids.

Schema: knowledge_sources, knowledge_chunks, knowledge_embeddings, and an FTS5
virtual table attached to knowledge_chunks as a DDL hook so it is created and
dropped with the table it indexes. Migration 92. A pre-M7 database opens
unchanged and needs no sources to play.

Bundle: the source content and the reader's judgements about it travel; the
passages, index rows and vectors are rebuilt on import, so a restored campaign
is searchable immediately without a reindex step.

One runtime dependency: python-multipart, Starlette's multipart parser. It is
what makes the upload surface possible, and the upload surface is why no
pathname is ever accepted.

The test doubles were the reason the defect shipped, so they were corrected too.
The retrieval stub scored unrelated text at 0.06-0.20 where the real model
scores it at 0.43-0.44, and its docstring said it had deliberately removed the
constant component that "would put a similarity floor under every pair" — which
is exactly the property real models have. The stub now has that floor, one test
fails if it is ever removed, and another reproduces the superseded rule and
asserts it is still fooled by the same fixture. Run against the pre-corrective
implementation, the new suite fails 13 of 18.

Tests: 939 passed, 14 skipped (836/7 at M6). 110 new across seven files, one of
which mocks nothing between itself and Ollama and re-measures the similarity
separation on every run. 43/43 checks in a real Firefox, reproduced.
Docker build clean.

Four other defects found by review or by the browser run were fixed here rather
than carried: an unreachable relevance constant that appeared to enforce
something and did not; acceptance tests using the wrong fixture files, so G07's
trap was never exercised; a bidirectional override surviving into displayed
filenames; and, from the implementation pass, the Insights panel showing M5's
two state sections as raw keys and the source inspector refetching on every
keystroke.

M7 was independently reviewed, which returned PASS WITH CORRECTIVE WORK
REQUIRED. Both blocking findings are closed, and closeout resolved the
embedding-model calibration boundary the corrective pass had left as debt.
planning/reports/M7-IMPLEMENTATION-REPORT.md carries the review, the corrective
closeout and the closeout verification in sequence, none overwriting another.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017HdaXiFbscatQaLS7dJk6b

backend/tests/test_knowledge_performance.py

@@ -0,0 +1,255 @@
+"""M7: the knowledge read paths must not grow a query per source or per passage.
+
+The same discipline `test_context_performance.py` holds for M6, applied to the
+four paths M7 adds. Each of them lists or joins over rows that a real library
+has many of, and each could plausibly have been written one query at a time:
+
+    source list           a chunk count and an embedded count per row
+    source detail         the source, and its passages
+    retrieval             lexical candidates, semantic candidates, their rows
+    context build         all of the above, inside a prompt assembly
+
+The assertions are on **growth**, not on an exact count: a fixed number breaks
+on any unrelated query and teaches the next person to raise it. What matters is
+that four times the library does not cost four times the queries.
+
+Also asserted here: candidates are bounded *in the database* before the Python
+reranking runs. "Do not load every chunk in the campaign merely to find the top
+few" is a statement about the SQL, so it is tested against the SQL.
+
+    python -m pytest tests/test_knowledge_performance.py -v
+"""
+
+import asyncio
+
+import pytest
+from fastapi import Depends
+from fastapi.testclient import TestClient
+from sqlalchemy import event, select
+
+from app import auth, limits, memorybank, models
+from app.database import Base, SessionLocal, engine, get_db
+from app.knowledge import embeddings, retrieval
+from app.main import app
+from app.routers import adventures
+
+from fakes import ScriptedProvider, state_block
+
+
+class StubEmbedder:
+    async def embed(self, texts):
+        return [[1.0, float(len(t) % 5), 0.5] for t in texts]
+
+
+@pytest.fixture()
+def sql_log():
+    statements: list[str] = []
+
+    def record(conn, cursor, statement, parameters, context, executemany):
+        statements.append(statement)
+
+    event.listen(engine, "before_cursor_execute", record)
+    try:
+        yield statements
+    finally:
+        event.remove(engine, "before_cursor_execute", record)
+
+
+@pytest.fixture()
+def client(monkeypatch):
+    Base.metadata.create_all(bind=engine)
+    memorybank._vector_cache.clear()
+    embeddings._cache.clear()
+    setup = SessionLocal()
+    user = models.User(is_guest=False, email="m7perf@example.com")
+    setup.add(user)
+    setup.flush()
+    setup.add(models.Settings(
+        user_id=user.id, model="test-model", embedding_model="nomic-embed-text",
+        context_token_budget=8000, max_output_tokens=400,
+    ))
+    adventure = models.Adventure(user_id=user.id, title="Performance")
+    setup.add(adventure)
+    setup.flush()
+    setup.add(models.Action(adventure_id=adventure.id, type="start",
+                            text="Aldric stands in the crypt beneath the Old Abbey."))
+    setup.commit()
+    adv_id, user_id = adventure.id, user.id
+    setup.close()
+
+    monkeypatch.setattr(limits, "check_row_cap", lambda *a, **k: None)
+    monkeypatch.setattr(adventures.turns, "OpenAICompatibleProvider", ScriptedProvider)
+    monkeypatch.setattr(memorybank, "embedding_provider", lambda s: StubEmbedder())
+    monkeypatch.setattr(memorybank, "summary_provider", lambda s: StubEmbedder())
+    app.dependency_overrides[auth.get_current_user] = (
+        lambda db=Depends(get_db): db.get(models.User, user_id)
+    )
+    test_client = TestClient(app)
+    test_client.adv_id = adv_id
+    test_client.user_id = user_id
+    try:
+        yield test_client
+    finally:
+        app.dependency_overrides.clear()
+        memorybank._vector_cache.clear()
+        embeddings._cache.clear()
+        Base.metadata.drop_all(bind=engine)
+
+
+def add_sources(client, count, paragraphs=6, prefix="lore"):
+    """Imports `count` sources, each with several passages of crypt-ish prose."""
+    for n in range(count):
+        body = "\n\n".join(
+            f"## {prefix} {n} section {p}\n\n"
+            + ("The crypt beneath the Old Abbey at Westhaven is vaulted in "
+               "stone, and the stair descends past niches cut for the dead. ") * 8
+            for p in range(paragraphs)
+        )
+        response = client.post(
+            f"/api/adventures/{client.adv_id}/knowledge",
+            files={"file": (f"{prefix}-{n}.md", body.encode(), "text/markdown")},
+            data={"classification": ["canon", "reference", "inspiration"][n % 3],
+                  "allow_duplicate": "true"},
+        )
+        assert response.status_code == 201, response.text[:200]
+
+
+def counts(client):
+    with SessionLocal() as db:
+        sources = len(db.execute(select(models.KnowledgeSource)).scalars().all())
+        chunks = len(db.execute(select(models.KnowledgeChunk)).scalars().all())
+    return sources, chunks
+
+
+def measure(sql_log, call):
+    sql_log.clear()
+    result = call()
+    return len(sql_log), result
+
+
+def retrieve(client):
+    with SessionLocal() as db:
+        adventure = db.get(models.Adventure, client.adv_id)
+        settings = db.execute(select(models.Settings).where(
+            models.Settings.user_id == client.user_id)).scalars().first()
+        return asyncio.run(retrieval.retrieve(adventure, settings))
+
+
+# --------------------------------------------------------------------- tests
+
+def test_the_source_list_does_not_cost_a_query_per_source(client, sql_log):
+    add_sources(client, 4)
+    small, _ = measure(sql_log, lambda: client.get(
+        f"/api/adventures/{client.adv_id}/knowledge").json())
+
+    add_sources(client, 12, prefix="more")
+    large, rows = measure(sql_log, lambda: client.get(
+        f"/api/adventures/{client.adv_id}/knowledge").json())
+
+    assert len(rows) == 16
+    assert large == small, f"{small} queries for 4 sources, {large} for 16"
+    # ...and the counts it shows are real, so the fixed query count is not
+    # because the counts were dropped.
+    assert all(row["chunk_count"] > 0 for row in rows)
+
+
+def test_source_detail_does_not_cost_a_query_per_passage(client, sql_log):
+    add_sources(client, 1, paragraphs=3)
+    small_id = client.get(f"/api/adventures/{client.adv_id}/knowledge").json()[0]["id"]
+    small, _ = measure(sql_log, lambda: client.get(
+        f"/api/adventures/{client.adv_id}/knowledge/{small_id}/chunks").json())
+
+    add_sources(client, 1, paragraphs=24, prefix="big")
+    big_id = client.get(f"/api/adventures/{client.adv_id}/knowledge").json()[-1]["id"]
+    large, chunks = measure(sql_log, lambda: client.get(
+        f"/api/adventures/{client.adv_id}/knowledge/{big_id}/chunks").json())
+
+    assert len(chunks) > 3
+    assert large == small, f"{small} queries for a small source, {large} for a big one"
+
+
+def test_retrieval_does_not_grow_with_the_library(client, sql_log):
+    add_sources(client, 4)
+    embed_pending(client)
+    small, small_result = measure(sql_log, lambda: retrieve(client))
+
+    add_sources(client, 16, prefix="more")
+    embed_pending(client)
+    embeddings.forget_cached(client.adv_id)
+    large, large_result = measure(sql_log, lambda: retrieve(client))
+
+    sources, chunks = counts(client)
+    assert sources == 20 and chunks > 40
+    assert small_result.candidates and large_result.candidates
+    assert large <= small + 1, f"{small} queries at 4 sources, {large} at 20"
+
+
+def test_the_context_build_does_not_grow_with_the_library(client, sql_log):
+    add_sources(client, 4)
+    embed_pending(client)
+    ScriptedProvider.replies = [f"The crypt is cold.\n{state_block([])}"]
+    client.post(f"/api/adventures/{client.adv_id}/actions",
+                json={"type": "do", "text": "Aldric descends into the crypt."})
+    small, _ = measure(sql_log, lambda: client.get(
+        f"/api/adventures/{client.adv_id}/context").json())
+
+    add_sources(client, 16, prefix="more")
+    embed_pending(client)
+    embeddings.forget_cached(client.adv_id)
+    large, report = measure(sql_log, lambda: client.get(
+        f"/api/adventures/{client.adv_id}/context").json())
+
+    assert report["knowledge"]["used"]
+    assert large <= small + 1, f"{small} queries at 4 sources, {large} at 20"
+
+
+def test_candidates_are_bounded_in_sql_before_the_python_ranking(client, sql_log):
+    """"Do not load every chunk merely to find the top few", asserted on the SQL."""
+    add_sources(client, 20, paragraphs=8)
+    embed_pending(client)
+    embeddings.forget_cached(client.adv_id)
+    _sources, chunks = counts(client)
+    assert chunks > retrieval.LEXICAL_CANDIDATES * 2, chunks
+
+    sql_log.clear()
+    result = retrieve(client)
+
+    # The lexical query names a LIMIT, and the merged candidate set is bounded
+    # by the two per-path caps rather than by the size of the library.
+    lexical = [s for s in sql_log if "knowledge_fts" in s and "MATCH" in s]
+    assert lexical, sql_log
+    assert all("LIMIT" in s for s in lexical)
+    assert result.considered <= (
+        retrieval.LEXICAL_CANDIDATES + retrieval.SEMANTIC_CANDIDATES
+    )
+    assert result.considered < chunks, (result.considered, chunks)
+
+    # The row fetch for those candidates is one query, not one per candidate.
+    loads = [s for s in sql_log
+             if "knowledge_chunks" in s and "knowledge_sources" in s
+             and " IN " in s.upper()]
+    assert len(loads) <= 2, loads
+
+
+def test_the_semantic_scan_reads_only_narrow_columns(client, sql_log):
+    """A vector is 6 kB; the catalogue read must not fetch passage text."""
+    add_sources(client, 6)
+    embed_pending(client)
+    embeddings.forget_cached(client.adv_id)
+
+    sql_log.clear()
+    retrieve(client)
+    catalogue = [s for s in sql_log
+                 if "knowledge_embeddings.chunk_id" in s
+                 and "knowledge_embeddings.vector" not in s]
+    assert catalogue, "the semantic catalogue read was not found"
+    assert all("knowledge_chunks.text" not in s for s in catalogue)
+
+
+def embed_pending(client):
+    with SessionLocal() as db:
+        adventure = db.get(models.Adventure, client.adv_id)
+        settings = db.execute(select(models.Settings).where(
+            models.Settings.user_id == client.user_id)).scalars().first()
+        asyncio.run(embeddings.embed_pending(db, adventure, settings))
+        db.commit()