Finding a thing that does not use your words

Three pieces, and the first one is that they are all optional.

Extraction stops being constants. Upload size, image edge, JPEG quality, PDF
pages, extracted characters, orphan age and the text-extension list are settings
now, read through a process-level snapshot rather than a session -- `prepare` and
everything under it are called from routes, tool runners and the startup sweep,
and several of those have no session in hand. Two things deliberately stayed
constants: the decompression-bomb guard, which is a guard and not a preference,
and ORPHAN_AGE, which would have been evaluated at import if it stayed in the
signature and pinned the shipped 24 hours whatever anybody set.

An embedding model is picked from the models an administrator flagged for it, and
one that has since lost its flag is *named* rather than dropped from the picker:
a setting that vanishes is one nobody can tell from a setting never made. Nothing
here is required. Choosing none means no chunk rows, no requests, and
retrieval.search returning exactly what fts.search_ids returns in exactly that
order -- asserted, because it is what makes this safe to land on an instance that
never asked for it.

The two rankings are fused by reciprocal rank fusion: ranks and not scores,
because bm25 is a corpus-dependent negative and cosine is 0..1, and normalising
them onto one scale means picking a constant nobody can tune without a labelled
set they do not have. RRF's one constant is famously insensitive and degrades to
whichever list is non-empty -- which is what turns "no embedding model" into a
branch that does not exist.

A record scores as its best chunk rather than its average, or a long document
about something else outranks a short one that says the thing. Width and model
are stored beside every vector and a mismatch is skipped, because vectors from
two spaces score against each other perfectly happily and mean nothing -- a
search that works and is wrong is the worst failure this can have, and a model
change now leaves stale rows ignored rather than trusted.

Indexing is fired and forgotten, and how a change is noticed is a session event
rather than a call in each of the ten library writers. That is a departure from
this codebase's taste for explicit seams, for the reason tool_label is a Jinja
global: a step every writer has to remember is one that gets forgotten, and here
forgetting is silent -- the record saves, keyword search still finds it, and only
its recall goes stale. Chunks are embedded before anything is deleted, so a
failure leaves the old index rather than half a new one.

Also: `embeddings` joins the model capabilities, and the three tool flags that
had shipped with no checkbox -- canvas, scheduling and helpers -- have one.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

src/lembas/services/library/chunks.py

@@ -0,0 +1,125 @@
+"""Splitting a record into pieces small enough to embed, and packing vectors.
+
+One implementation, used by documents, notes, skills and reports. Three would
+drift, and drift here is invisible: a splitter that behaves differently for
+notes than for documents produces a search that works and ranks wrongly.
+
+## How it splits
+
+On **paragraph boundaries first**, falling back to lines and then to a hard cut,
+because a chunk that ends mid-sentence is one whose embedding is about half a
+thought. The overlap carries the tail of the previous chunk into the next, so a
+sentence that straddles a boundary is whole in one of them.
+
+Characters rather than tokens throughout. The count has to be made without
+asking the endpoint -- `services/tokens.py` already establishes four characters
+to a token as this codebase's estimate, and being 20% out about a chunk size is
+a slightly different chunk, not a wrong one.
+
+## Packing
+
+float32, little-endian. A 1024-dimension vector is 4KB packed and about 20KB as
+JSON text, and every one of them is read on every semantic search.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import struct
+
+# Below this a piece is not worth a row: the embedding of six words is mostly
+# noise, and a search that returns "and the following:" as its best hit is worse
+# than one that returns nothing.
+MIN_CHUNK_CHARS = 40
+
+
+def split(text: str, *, size: int = 1200, overlap: int = 150) -> list[str]:
+    """A record's text as pieces of roughly `size` characters.
+
+    `overlap` is how much of the previous piece rides along with the next. It is
+    clamped to half the size here as well as in the settings accessor, because
+    an overlap at or past the size means every piece starts where the last one
+    did and the loop never advances -- a hang rather than a bad index, so it is
+    refused in both places rather than in the more convenient one.
+    """
+    body = (text or "").strip()
+    if not body:
+        return []
+    size = max(200, int(size))
+    overlap = max(0, min(int(overlap), size // 2))
+    if len(body) <= size:
+        return [body]
+
+    pieces: list[str] = []
+    start = 0
+    while start < len(body):
+        end = min(start + size, len(body))
+        if end < len(body):
+            end = _boundary(body, start, end)
+        piece = body[start:end].strip()
+        if len(piece) >= MIN_CHUNK_CHARS:
+            pieces.append(piece)
+        if end >= len(body):
+            break
+        start = max(end - overlap, start + 1)
+    return pieces
+
+
+def _boundary(body: str, start: int, end: int) -> int:
+    """Where to cut, preferring a paragraph break and then a line break.
+
+    Searched backwards from the hard limit, and only within the last third of
+    the piece: a paragraph break near the *start* would produce a chunk a
+    fraction of the size, which is how a long document turns into hundreds of
+    tiny rows that each match nothing.
+    """
+    floor = start + (end - start) * 2 // 3
+    for marker in ("\n\n", "\n", ". "):
+        found = body.rfind(marker, floor, end)
+        if found > floor:
+            return found + len(marker)
+    return end
+
+
+def digest(text: str) -> str:
+    """A hash of what a chunk set was built from.
+
+    What makes re-indexing an unchanged record free, and what makes "is this
+    index current?" answerable without embedding anything. sha256 rather than
+    md5 for no reason beyond having no reason to prefer md5; both are being used
+    as a change detector rather than against an adversary.
+    """
+    return hashlib.sha256((text or "").encode("utf-8")).hexdigest()
+
+
+def pack(vector: list[float]) -> bytes:
+    return struct.pack(f"<{len(vector)}f", *vector)
+
+
+def unpack(blob: bytes, dims: int) -> list[float]:
+    """A stored vector, or an empty list if the row does not add up.
+
+    Length is checked against the declared width rather than inferred from it: a
+    truncated BLOB would otherwise unpack into a shorter vector and score
+    against a query happily, which is a wrong answer rather than a missing one.
+    """
+    if dims <= 0 or len(blob) != dims * 4:
+        return []
+    return list(struct.unpack(f"<{dims}f", blob))
+
+
+def dot(left: list[float], right: list[float]) -> float:
+    """Cosine similarity, given that both sides are already unit vectors.
+
+    Normalisation happens once, at write time, in `llm/embeddings.py` -- so
+    every comparison here is a multiply-and-add rather than two square roots per
+    pair. A width mismatch scores zero rather than raising: it means the vectors
+    came from two different models, and the honest answer to "how similar are
+    these?" across two spaces is "this tells you nothing".
+    """
+    if len(left) != len(right) or not left:
+        return 0.0
+    return sum(a * b for a, b in zip(left, right, strict=True))
+
+
+__all__ = ["MIN_CHUNK_CHARS", "digest", "dot", "pack", "split", "unpack"]