danvics/haiku.rag
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Time the search pathway inside an agent run

Browse source 
The pydantic-ai instrumentation timed the rag_search tool call as a
whole; everything between the tool boundary and LanceDB was one opaque
block. Decompose it into embed / execute / hydrate / rerank / expand /
images, and attribute the two one-off costs (store open, reranker
weight load) that a run's first tool call would otherwise absorb
silently.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
bryan davis 2026-08-13 14:33:06 -05:00
parent d864cf8008
commit 38446f9618
No known key found for this signature in database
GPG key ID: D11B4A4C0C731E5E
5 changed files with 665 additions and 124 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

52
haiku_rag_slim/haiku/rag/capabilities/_base.py
View file

@ -33,6 +33,7 @@ from haiku.rag.client import HaikuRAG
from haiku.rag.config.models import AppConfig from haiku.rag.config.models import AppConfig
from haiku.rag.store.models.chunk import SearchResult from haiku.rag.store.models.chunk import SearchResult
from haiku.rag.store.models.citation import Citation, resolve_citations from haiku.rag.store.models.citation import Citation, resolve_citations
from haiku.rag.telemetry import logfire
from haiku.rag.tools.search import build_image_content_from_results from haiku.rag.tools.search import build_image_content_from_results
CITATION_GRACE_REQUESTS = 2 CITATION_GRACE_REQUESTS = 2
@ -350,9 +351,13 @@ class RAGCapabilityBase[StateT: BaseModel](AbstractCapability[Any]):
if self.rag is None: if self.rag is None:
async with self.resource_lock: async with self.resource_lock:
if self.rag is None: if self.rag is None:
rag = HaikuRAG(self.db_path, config=self.config, read_only=True) # Opening the store is lazy, so the run's first tool call
await rag.__aenter__() # pays it. Spanned separately or it reads as "the first
self.rag = rag # search was slow" with nothing to point at.
with logfire.span("rag.client.open", db_path=str(self.db_path)):
rag = HaikuRAG(self.db_path, config=self.config, read_only=True)
await rag.__aenter__()
self.rag = rag
return self.rag return self.rag
async def get_picture_bytes(self, document_id: str, self_ref: str) -> bytes | None: async def get_picture_bytes(self, document_id: str, self_ref: str) -> bytes | None:
@ -430,19 +435,34 @@ class RAGCapabilityBase[StateT: BaseModel](AbstractCapability[Any]):
"Search limit reached. Answer the question using " "Search limit reached. Answer the question using "
"the results you already have." "the results you already have."
) )
async with self.rag_lock: # Opened after the budget check so the span means "a search ran" —
formatted, results = await search_corpus( # a refused call is already visible as the tool span's exception.
await self._ensure_rag(), # `search_index` is what shows whether cost grows across a run or
query, # one search was pathological; the tool span alone cannot tell them
limit=limit, # apart, and the first search also carries the one-off store open.
document_filter=getattr(self.state, "document_filter", None), with logfire.span(
) "ask.tool.search",
state = cast(Any, self.state) namespace=self.state_namespace,
state.searches[query] = results search_index=self.search_count,
self._note_evidence() max_searches=self._max_searches,
if self.vision and (parts := build_image_content_from_results(results)): limit=limit,
return ToolReturn(return_value=formatted, content=parts) ) as span:
return formatted async with self.rag_lock:
formatted, results = await search_corpus(
await self._ensure_rag(),
query,
limit=limit,
document_filter=getattr(self.state, "document_filter", None),
)
span.set_attribute("results", len(results))
span.set_attribute("formatted_chars", len(formatted))
state = cast(Any, self.state)
state.searches[query] = results
self._note_evidence()
if self.vision and (parts := build_image_content_from_results(results)):
span.set_attribute("image_parts", len(parts))
return ToolReturn(return_value=formatted, content=parts)
return formatted
async def _cite(self, chunk_ids: list[str]) -> str: async def _cite(self, chunk_ids: list[str]) -> str:
"""Register the evidence behind this answer, or declare there is none. """Register the evidence behind this answer, or declare there is none.

169
haiku_rag_slim/haiku/rag/client/search.py
View file

@ -4,11 +4,34 @@ from typing import TYPE_CHECKING
from haiku.rag.store.models.chunk import Chunk, SearchResult, SearchType from haiku.rag.store.models.chunk import Chunk, SearchResult, SearchType
from haiku.rag.store.models.document_item import PICTURE_REF_PREFIX from haiku.rag.store.models.document_item import PICTURE_REF_PREFIX
from haiku.rag.telemetry import logfire
if TYPE_CHECKING: if TYPE_CHECKING:
from PIL import Image as PILImage from PIL import Image as PILImage
from haiku.rag.client import HaikuRAG from haiku.rag.client import HaikuRAG
from haiku.rag.reranking.base import RerankerBase
def _get_reranker(client: "HaikuRAG") -> "RerankerBase | None":
"""Materialize the client's reranker, timing the first touch.
``HaikuRAG.reranker`` is a ``cached_property``, and the local rerankers
load model weights in their constructor — seconds of synchronous work on
the event loop that otherwise lands unattributed inside whichever search
happened to be first. Every later search hits the cache and emits
nothing, so a ``search.reranker.load`` span in a trace means a cold
process, not a slow reranker.
"""
if "reranker" in client.__dict__:
return client.reranker
model = client._config.reranking.model
with logfire.span(
"search.reranker.load",
provider=model.provider if model else None,
model=model.name if model else None,
):
return client.reranker
async def search( async def search(
@ -42,7 +65,7 @@ async def search(
if search_type is None: if search_type is None:
search_type = "hybrid" search_type = "hybrid"
reranker = client.reranker reranker = _get_reranker(client)
if reranker is None: if reranker is None:
chunk_results = await client.chunk_repository.search( chunk_results = await client.chunk_repository.search(
@ -149,51 +172,64 @@ async def _populate_image_data(client: "HaikuRAG", results: list[SearchResult])
if r.document_id and r.doc_item_refs: if r.document_id and r.doc_item_refs:
by_doc.setdefault(r.document_id, []).append(r) by_doc.setdefault(r.document_id, []).append(r)
for doc_id, doc_results in by_doc.items(): # Up to three reads per document plus base64 encoding of every blob, so
all_refs = {ref for r in doc_results for ref in r.doc_item_refs} # `documents` is the round-trip multiplier and `bytes` the encoding load.
caption_to_picture = await repo.get_caption_picture_refs(doc_id, list(all_refs)) # Both are counted raw (pre-base64); the encoded payload is ~4/3 of it.
with logfire.span("search.images", documents=len(by_doc)) as span:
picture_count = 0
picture_bytes = 0
for doc_id, doc_results in by_doc.items():
all_refs = {ref for r in doc_results for ref in r.doc_item_refs}
caption_to_picture = await repo.get_caption_picture_refs(
doc_id, list(all_refs)
)
result_pictures: list[tuple[SearchResult, list[str]]] = [] result_pictures: list[tuple[SearchResult, list[str]]] = []
wanted: list[str] = [] wanted: list[str] = []
seen: set[str] = set() seen: set[str] = set()
for r in doc_results: for r in doc_results:
pictures: list[str] = [] pictures: list[str] = []
for ref in r.doc_item_refs: for ref in r.doc_item_refs:
picture = ( picture = (
ref ref
if ref.startswith(PICTURE_REF_PREFIX) if ref.startswith(PICTURE_REF_PREFIX)
else caption_to_picture.get(ref) else caption_to_picture.get(ref)
) )
if picture and picture not in pictures: if picture and picture not in pictures:
pictures.append(picture) pictures.append(picture)
if pictures: if pictures:
result_pictures.append((r, pictures)) result_pictures.append((r, pictures))
for picture in pictures: for picture in pictures:
if picture not in seen: if picture not in seen:
wanted.append(picture) wanted.append(picture)
seen.add(picture) seen.add(picture)
if not wanted: if not wanted:
continue continue
bytes_by_ref = await repo.get_pictures_for_chunk(doc_id, wanted) bytes_by_ref = await repo.get_pictures_for_chunk(doc_id, wanted)
if not bytes_by_ref: if not bytes_by_ref:
continue continue
captions_by_ref = await repo.get_text_for_refs( captions_by_ref = await repo.get_text_for_refs(
doc_id, list(bytes_by_ref.keys()) doc_id, list(bytes_by_ref.keys())
) )
for r, pictures in result_pictures: for r, pictures in result_pictures:
attached: dict[str, str] = {} attached: dict[str, str] = {}
captions: dict[str, str] = {} captions: dict[str, str] = {}
for ref in pictures: for ref in pictures:
blob = bytes_by_ref.get(ref) blob = bytes_by_ref.get(ref)
if blob: if blob:
attached[ref] = base64.b64encode(blob).decode("ascii") attached[ref] = base64.b64encode(blob).decode("ascii")
caption = captions_by_ref.get(ref) picture_count += 1
if caption: picture_bytes += len(blob)
captions[ref] = caption caption = captions_by_ref.get(ref)
if attached: if caption:
r.image_data = attached captions[ref] = caption
if captions: if attached:
r.picture_captions = captions r.image_data = attached
if captions:
r.picture_captions = captions
span.set_attribute("pictures", picture_count)
span.set_attribute("bytes", picture_bytes)
async def expand_context( async def expand_context(
@ -223,23 +259,38 @@ async def expand_context(
expanded_results = [] expanded_results = []
for doc_id, doc_results in document_groups.items(): # Each expanded document costs two round trips (resolve_refs, then
if doc_id is None: # get_items_in_range), so `documents` is the multiplier on this stage.
expanded_results.extend(doc_results) # `context_chars` is the payload the model is about to be handed, which
continue # ties retrieval time to the next request's prompt size.
with logfire.span(
"search.expand",
documents=len(document_groups),
max_chars=max_chars,
results_in=len(search_results),
) as span:
for doc_id, doc_results in document_groups.items():
if doc_id is None:
expanded_results.extend(doc_results)
continue
has_refs = any(r.doc_item_refs for r in doc_results) has_refs = any(r.doc_item_refs for r in doc_results)
if not has_refs: if not has_refs:
expanded_results.extend(doc_results) expanded_results.extend(doc_results)
continue continue
expanded = await expand_with_items( expanded = await expand_with_items(
client.document_item_repository, client.document_item_repository,
doc_id, doc_id,
doc_results, doc_results,
max_chars, max_chars,
)
expanded_results.extend(expanded)
span.set_attribute("results_out", len(expanded_results))
span.set_attribute(
"context_chars", sum(len(r.content) for r in expanded_results)
) )
expanded_results.extend(expanded)
expanded_results.sort(key=lambda r: r.score, reverse=True) expanded_results.sort(key=lambda r: r.score, reverse=True)
# image_data and picture_captions are preserved through expansion by # image_data and picture_captions are preserved through expansion by

23
haiku_rag_slim/haiku/rag/reranking/base.py
View file

@ -1,5 +1,6 @@
from haiku.rag.config import Config from haiku.rag.config import Config
from haiku.rag.store.models.chunk import Chunk from haiku.rag.store.models.chunk import Chunk
from haiku.rag.telemetry import logfire
class RerankerBase: class RerankerBase:
@ -10,7 +11,27 @@ class RerankerBase:
) -> list[tuple[Chunk, float]]: ) -> list[tuple[Chunk, float]]:
if not chunks: if not chunks:
return [] return []
return await self._rerank(query, chunks, top_n) # Instrumented here rather than in each subclass: this is the one
# choke point every provider passes through, so the span covers
# local weights and remote APIs alike. `candidates` is the search's
# limit*10 fan-out, the main lever on how long scoring takes.
with logfire.span(
"search.rerank",
provider=type(self).__name__,
model=self._model,
candidates=len(chunks),
top_n=top_n,
) as span:
results = await self._rerank(query, chunks, top_n)
scores = [score for _, score in results]
span.set_attribute("results", len(results))
# Spread separates a reranker that discriminated from one whose
# scores saturated onto the same value, leaving order to the sort.
span.set_attribute("top_score", max(scores) if scores else None)
span.set_attribute(
"score_spread", max(scores) - min(scores) if scores else None
)
return results
async def _rerank( async def _rerank(
self, query: str, chunks: list[Chunk], top_n: int = 10 self, query: str, chunks: list[Chunk], top_n: int = 10

129
haiku_rag_slim/haiku/rag/store/repositories/chunk.py
View file

@ -11,6 +11,7 @@ from lancedb.rerankers import RRFReranker
from haiku.rag.store.engine import Store, query_to_pydantic from haiku.rag.store.engine import Store, query_to_pydantic
from haiku.rag.store.models.chunk import Chunk, SearchType from haiku.rag.store.models.chunk import Chunk, SearchType
from haiku.rag.telemetry import logfire
from haiku.rag.utils import escape_sql_string from haiku.rag.utils import escape_sql_string
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@ -205,6 +206,23 @@ class ChunkRepository:
await self.store.chunks_table.delete(f"document_id = '{document_id}'") await self.store.chunks_table.delete(f"document_id = '{document_id}'")
return True return True
async def _embed_query(self, query: str) -> list[float]:
"""Embed a search query, timed as its own stage.
Instrumented here rather than in ``EmbedderWrapper.embed_query``
because the multimodal embedders (vllm, cohere, voyageai) override
that method — the very providers whose embedding is a remote round
trip. Wrapping the call site covers every embedder uniformly.
"""
with logfire.span(
"search.embed",
provider=self.store._config.embeddings.model.provider,
model=self.store._config.embeddings.model.name,
) as span:
embedding = await self.embedder.embed_query(query)
span.set_attribute("dim", len(embedding))
return embedding
async def search( async def search(
self, self,
query: str = "", query: str = "",
@ -255,7 +273,7 @@ class ChunkRepository:
.refine_factor(self.store._config.search.vector_refine_factor) .refine_factor(self.store._config.search.vector_refine_factor)
) )
elif search_type == "vector": elif search_type == "vector":
query_embedding = await self.embedder.embed_query(query) query_embedding = await self._embed_query(query)
results = ( results = (
self.store.chunks_table.query() self.store.chunks_table.query()
.nearest_to(query_embedding) .nearest_to(query_embedding)
@ -267,7 +285,7 @@ class ChunkRepository:
query, columns="content_fts" query, columns="content_fts"
) )
else: # hybrid (default) else: # hybrid (default)
query_embedding = await self.embedder.embed_query(query) query_embedding = await self._embed_query(query)
reranker = RRFReranker() reranker = RRFReranker()
results = ( results = (
self.store.chunks_table.query() self.store.chunks_table.query()
@ -402,55 +420,70 @@ class ChunkRepository:
else: else:
raise ValueError("Unknown search result format, cannot extract scores") raise ValueError("Unknown search result format, cannot extract scores")
df = await query_result.to_pandas() # The query builder is lazy — `nearest_to`/`nearest_to_text`/`rerank`
# only describe the query. This await is where LanceDB actually runs
# the ANN and FTS passes, so it is the stage worth timing.
with logfire.span("search.execute") as span:
df = await query_result.to_pandas()
span.set_attribute("rows", len(df))
# Extract scores # Turning the frame into Chunks costs one more LanceDB read (the
scores = extract_scores(df) # document metadata batch) plus a per-row json.loads — at the
# limit*10 candidate fan-out a reranked search uses, that is not
# free. Timed apart from `search.execute` so a slow search can be
# blamed on LanceDB or on this, but not ambiguously on both.
with logfire.span("search.hydrate") as span:
# Extract scores
scores = extract_scores(df)
# Convert DataFrame rows to ChunkRecords # Convert DataFrame rows to ChunkRecords
pydantic_results = [ pydantic_results = [
self.store.ChunkRecord( self.store.ChunkRecord(
id=str(row["id"]), id=str(row["id"]),
document_id=str(row["document_id"]), document_id=str(row["document_id"]),
content=str(row["content"]), content=str(row["content"]),
content_fts=str(row.get("content_fts", "")), content_fts=str(row.get("content_fts", "")),
metadata=str(row["metadata"]), metadata=str(row["metadata"]),
order=int(row["order"]) if "order" in row else 0, order=int(row["order"]) if "order" in row else 0,
) )
for _, row in df.iterrows() for _, row in df.iterrows()
] ]
# Collect all unique document IDs for batch lookup # Collect all unique document IDs for batch lookup
document_ids = list(set(chunk.document_id for chunk in pydantic_results)) document_ids = list(set(chunk.document_id for chunk in pydantic_results))
span.set_attribute("documents", len(document_ids))
# Batch fetch document metadata (skip content/docling blobs) # Batch fetch document metadata (skip content/docling blobs)
documents_map: dict[str, dict] = {} documents_map: dict[str, dict] = {}
if document_ids: if document_ids:
id_list = "', '".join(document_ids) id_list = "', '".join(document_ids)
where_clause = f"id IN ('{id_list}')" where_clause = f"id IN ('{id_list}')"
doc_rows = await ( doc_rows = await (
self.store.document_meta_table.query() self.store.document_meta_table.query()
.select(["id", "uri", "title", "metadata"]) .select(["id", "uri", "title", "metadata"])
.where(where_clause) .where(where_clause)
.to_list() .to_list()
) )
documents_map = {str(row["id"]): row for row in doc_rows} documents_map = {str(row["id"]): row for row in doc_rows}
# Build final results with document info # Build final results with document info
chunks_with_scores = [] chunks_with_scores = []
for i, chunk_record in enumerate(pydantic_results): for i, chunk_record in enumerate(pydantic_results):
doc = documents_map.get(chunk_record.document_id) doc = documents_map.get(chunk_record.document_id)
chunk = Chunk( chunk = Chunk(
id=chunk_record.id, id=chunk_record.id,
document_id=chunk_record.document_id, document_id=chunk_record.document_id,
content=chunk_record.content, content=chunk_record.content,
metadata=json.loads(chunk_record.metadata), metadata=json.loads(chunk_record.metadata),
order=chunk_record.order, order=chunk_record.order,
document_uri=doc["uri"] if doc else None, document_uri=doc["uri"] if doc else None,
document_title=doc["title"] if doc else None, document_title=doc["title"] if doc else None,
document_meta=json.loads(doc.get("metadata", "{}") if doc else "{}"), document_meta=json.loads(
) doc.get("metadata", "{}") if doc else "{}"
score = scores[i] if i < len(scores) else 1.0 ),
chunks_with_scores.append((chunk, score)) )
score = scores[i] if i < len(scores) else 1.0
chunks_with_scores.append((chunk, score))
return chunks_with_scores span.set_attribute("chunks", len(chunks_with_scores))
return chunks_with_scores

416
tests/test_search_telemetry.py Normal file
View file

@ -0,0 +1,416 @@
"""Span coverage for the search pathway.
The pydantic-ai instrumentation already times the ``rag_search`` tool call as
a whole; these spans decompose that duration into embed / execute / hydrate /
rerank / expand / images so a slow search can be attributed to a stage, plus
the two one-off costs (store open, reranker weight load) that the run's first
tool call would otherwise absorb silently. The tests assert the span names and
the attributes queries are written against — renaming either breaks saved
Logfire views, so they are pinned here.
Spans are captured by monkeypatching each module's ``logfire`` object, the
same approach ``tests/ingester/test_workers.py`` uses for breaker events.
"""
from contextlib import contextmanager
from dataclasses import dataclass, field
from functools import cached_property
from types import SimpleNamespace
import pandas as pd
import pytest
from haiku.rag.capabilities import _base as capabilities_base
from haiku.rag.client import search as search_module
from haiku.rag.config import Config
from haiku.rag.reranking import base as reranking_base
from haiku.rag.reranking.base import RerankerBase
from haiku.rag.store.models.chunk import Chunk, SearchResult
from haiku.rag.store.repositories import chunk as chunk_repo_module
from haiku.rag.store.repositories.chunk import ChunkRepository
@dataclass
class RecordedSpan:
name: str
attributes: dict = field(default_factory=dict)
def set_attribute(self, key, value):
self.attributes[key] = value
class SpanRecorder:
"""Stand-in for the scoped ``logfire`` instance that records spans."""
def __init__(self):
self.spans: list[RecordedSpan] = []
@contextmanager
def span(self, name, **attributes):
recorded = RecordedSpan(name, dict(attributes))
self.spans.append(recorded)
yield recorded
def by_name(self, name: str) -> RecordedSpan:
matches = [span for span in self.spans if span.name == name]
assert matches, f"no {name!r} span; recorded {[s.name for s in self.spans]}"
assert len(matches) == 1, f"expected one {name!r} span, got {len(matches)}"
return matches[0]
def names(self) -> list[str]:
return [span.name for span in self.spans]
@pytest.fixture
def recorder(monkeypatch):
"""Record spans from every module on the search path."""
recorder = SpanRecorder()
for module in (
reranking_base,
chunk_repo_module,
search_module,
capabilities_base,
):
monkeypatch.setattr(module, "logfire", recorder)
return recorder
class StubReranker(RerankerBase):
"""Reranker that returns fixed scores, so the span is tested without a
provider (local weights or remote API)."""
_model = "stub-reranker"
def __init__(self, scores: list[float]):
self._scores = scores
async def _rerank(self, query, chunks, top_n=10):
return list(zip(chunks, self._scores))[:top_n]
def _chunks(count: int) -> list[Chunk]:
return [Chunk(id=f"c{i}", content=f"chunk {i}") for i in range(count)]
async def test_rerank_span_records_fan_out_and_scores(recorder):
reranker = StubReranker([0.9, 0.5, 0.2])
await reranker.rerank("query", _chunks(3), top_n=2)
span = recorder.by_name("search.rerank")
assert span.attributes["provider"] == "StubReranker"
assert span.attributes["model"] == "stub-reranker"
# candidates is the limit*10 fan-out the search applied, the main lever
# on rerank latency; top_n is what the caller asked for.
assert span.attributes["candidates"] == 3
assert span.attributes["top_n"] == 2
assert span.attributes["results"] == 2
assert span.attributes["top_score"] == 0.9
assert span.attributes["score_spread"] == pytest.approx(0.4)
async def test_rerank_span_not_emitted_for_empty_candidates(recorder):
assert await StubReranker([]).rerank("query", [], top_n=5) == []
assert recorder.spans == []
async def test_rerank_span_tolerates_empty_results(recorder):
"""A reranker that drops every candidate still closes its span; the score
attributes go None rather than raising on max() of an empty sequence."""
await StubReranker([]).rerank("query", _chunks(2), top_n=2)
span = recorder.by_name("search.rerank")
assert span.attributes["results"] == 0
assert span.attributes["top_score"] is None
assert span.attributes["score_spread"] is None
class StubEmbedder:
async def embed_query(self, text):
return [0.1, 0.2, 0.3]
async def test_embed_span_records_provider_and_dim(recorder):
store = SimpleNamespace(embedder=StubEmbedder(), _config=Config)
repository = ChunkRepository(store) # type: ignore[arg-type]
assert await repository._embed_query("query") == [0.1, 0.2, 0.3]
span = recorder.by_name("search.embed")
assert span.attributes["provider"] == Config.embeddings.model.provider
assert span.attributes["model"] == Config.embeddings.model.name
assert span.attributes["dim"] == 3
class StubQueryResult:
"""LanceDB query builder stub: the awaited ``to_pandas`` is the point at
which the real builder stops being lazy and the search actually runs."""
def __init__(self, df: pd.DataFrame):
self._df = df
async def to_pandas(self):
return self._df
class StubTableQuery:
def __init__(self, rows: list[dict]):
self._rows = rows
def select(self, _columns):
return self
def where(self, _clause):
return self
async def to_list(self):
return self._rows
async def test_execute_span_records_row_count(recorder):
df = pd.DataFrame(
[
{
"id": "c0",
"document_id": "d0",
"content": "hello",
"metadata": "{}",
"order": 0,
"_relevance_score": 0.75,
}
]
)
doc_rows = [{"id": "d0", "uri": "file://d0", "title": None, "metadata": "{}"}]
store = SimpleNamespace(
embedder=StubEmbedder(),
_config=Config,
ChunkRecord=_chunk_record_type(),
document_meta_table=SimpleNamespace(query=lambda: StubTableQuery(doc_rows)),
)
repository = ChunkRepository(store) # type: ignore[arg-type]
results = await repository._process_search_results(StubQueryResult(df)) # type: ignore[arg-type]
assert len(results) == 1
# rows is the candidate count LanceDB returned, before rerank narrows it.
assert recorder.by_name("search.execute").attributes["rows"] == 1
# Turning that frame into Chunks is a separate stage: one more LanceDB
# read for document metadata plus a per-row json.loads.
hydrate = recorder.by_name("search.hydrate")
assert hydrate.attributes["documents"] == 1
assert hydrate.attributes["chunks"] == 1
# Ordering matters for reading a trace: execution precedes hydration.
assert recorder.names() == ["search.execute", "search.hydrate"]
def _chunk_record_type():
"""The record model Store builds per embedding dimension; only the fields
_process_search_results populates matter here."""
from pydantic import BaseModel
class ChunkRecord(BaseModel):
id: str
document_id: str
content: str
content_fts: str = ""
metadata: str = "{}"
order: int = 0
return ChunkRecord
async def test_expand_span_records_context_size(recorder):
"""Results without a document_id pass through unexpanded, so the span is
exercised without a document_items table behind it."""
client = SimpleNamespace(_config=Config, document_item_repository=None)
results = [
SearchResult(content="a" * 10, score=0.9),
SearchResult(content="b" * 5, score=0.4),
]
expanded = await search_module.expand_context(client, results) # type: ignore[arg-type]
assert len(expanded) == 2
span = recorder.by_name("search.expand")
assert span.attributes["documents"] == 1
assert span.attributes["max_chars"] == Config.search.max_context_chars
assert span.attributes["results_in"] == 2
assert span.attributes["results_out"] == 2
# context_chars is the payload about to be handed to the model, which ties
# this stage's cost to the next model request's prompt size.
assert span.attributes["context_chars"] == 15
class StubItemRepository:
"""Document-item repository serving one picture for one document."""
def __init__(self, picture_ref: str, blob: bytes):
self._picture_ref = picture_ref
self._blob = blob
async def get_caption_picture_refs(self, _document_id, _refs):
return {}
async def get_pictures_for_chunk(self, _document_id, _refs):
return {self._picture_ref: self._blob}
async def get_text_for_refs(self, _document_id, _refs):
return {self._picture_ref: "a caption"}
async def test_images_span_records_documents_and_bytes(recorder):
picture_ref = "#/pictures/0"
blob = b"\x89PNG" + b"x" * 60
client = SimpleNamespace(
_config=Config,
document_item_repository=StubItemRepository(picture_ref, blob),
)
results = [
SearchResult(
content="figure",
score=0.9,
document_id="d0",
doc_item_refs=[picture_ref],
)
]
await search_module._populate_image_data(client, results) # type: ignore[arg-type]
assert results[0].image_data is not None
span = recorder.by_name("search.images")
# documents is the round-trip multiplier — up to three reads each.
assert span.attributes["documents"] == 1
assert span.attributes["pictures"] == 1
# Raw blob bytes, not the ~4/3 larger base64 that gets attached.
assert span.attributes["bytes"] == len(blob)
async def test_images_span_records_zero_when_nothing_attaches(recorder):
"""A result set with no pictures still opens the span, so the absence of
image work is visible rather than inferred from a missing span."""
client = SimpleNamespace(_config=Config, document_item_repository=None)
await search_module._populate_image_data(
client, # type: ignore[arg-type]
[SearchResult(content="text only", score=0.5)],
)
span = recorder.by_name("search.images")
assert span.attributes["documents"] == 0
assert span.attributes["pictures"] == 0
assert span.attributes["bytes"] == 0
class StubRerankerClient:
"""Client whose ``reranker`` is a cached_property, like ``HaikuRAG``."""
def __init__(self, config):
self._config = config
self.loads = 0
@cached_property
def reranker(self):
self.loads += 1
return StubReranker([1.0])
async def test_reranker_load_span_emitted_once_per_process(recorder):
client = StubRerankerClient(Config)
first = search_module._get_reranker(client) # type: ignore[arg-type]
second = search_module._get_reranker(client) # type: ignore[arg-type]
# The cached_property loaded weights exactly once; the span marks that
# cold touch, so a search.reranker.load in a trace means a cold process.
assert first is second
assert client.loads == 1
assert recorder.names() == ["search.reranker.load"]
class StubRag:
"""Stands in for HaikuRAG so _ensure_rag opens no real store."""
opened = 0
def __init__(self, *_args, **_kwargs):
pass
async def __aenter__(self):
type(self).opened += 1
return self
async def __aexit__(self, *_exc):
return False
@pytest.fixture
def capability(tmp_path, monkeypatch):
from haiku.rag.capabilities.rag import RAGState, create_capability
monkeypatch.setattr(capabilities_base, "HaikuRAG", StubRag)
StubRag.opened = 0
built = create_capability(db_path=tmp_path / "test.lancedb", config=Config)
built.state = RAGState()
return built
async def test_client_open_span_emitted_once_per_run(recorder, capability):
first = await capability._ensure_rag()
second = await capability._ensure_rag()
# Store open is lazy and would otherwise be charged to whichever search
# ran first; the double-checked lock means only one span per run.
assert first is second
assert StubRag.opened == 1
span = recorder.by_name("rag.client.open")
assert span.attributes["db_path"] == str(capability.db_path)
async def test_tool_search_span_records_position_in_run(
recorder, capability, monkeypatch
):
results = [SearchResult(content="evidence", score=0.9, chunk_id="c0")]
async def _search_corpus(_rag, _query, limit=None, document_filter=None):
return "formatted evidence", results
monkeypatch.setattr(capabilities_base, "search_corpus", _search_corpus)
await capability._search("first query", limit=3)
await capability._search("second query", limit=3)
searches = [s for s in recorder.spans if s.name == "ask.tool.search"]
assert [s.attributes["search_index"] for s in searches] == [1, 2]
assert searches[0].attributes["namespace"] == "rag"
assert searches[0].attributes["max_searches"] == Config.qa.max_searches
assert searches[0].attributes["limit"] == 3
assert searches[0].attributes["results"] == 1
assert searches[0].attributes["formatted_chars"] == len("formatted evidence")
# Spans are recorded as they open, so this order is the nesting: the
# one-off store open happens inside the FIRST search and never again,
# which is what makes search #1 legitimately slower than the rest.
assert recorder.names() == [
"ask.tool.search",
"rag.client.open",
"ask.tool.search",
]
async def test_tool_search_span_not_emitted_when_budget_spent(
recorder, capability, monkeypatch
):
"""A refused call emits no span: the ToolFailed is already on the tool
span pydantic-ai opened, and search.* spans should mean work happened."""
from pydantic_ai import ToolFailed
async def _search_corpus(_rag, _query, limit=None, document_filter=None):
return "", []
monkeypatch.setattr(capabilities_base, "search_corpus", _search_corpus)
capability.search_count = capability._max_searches
with pytest.raises(ToolFailed):
await capability._search("one too many", limit=3)
assert "ask.tool.search" not in recorder.names()