Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: stop web results crowding attached files out of Deep Research (#17434)

Nilayce489c168684
diff --git a/studio/backend/core/research_runs.py b/studio/backend/core/research_runs.py
index b0dd90cc7f4..edffd42e7e9 100644
--- a/studio/backend/core/research_runs.py
+++ b/studio/backend/core/research_runs.py
@@ -713,6 +713,22 @@ def _fit_source_catalog(catalog: str, max_chars: int) -> str:
return ("".join(kept) if not kept or kept[0].endswith("\n") else "\n\n".join(kept)).rstrip()
+def _fit_synthesis_catalogs(
+ web_entries: list[str], document_entries: list[str], max_chars: int
+) -> tuple[str, str]:
+ limits = _water_fill([len("\n".join(web_entries)), len("\n".join(document_entries))], max_chars)
+ fitted = []
+ for entries, limit in zip((web_entries, document_entries), limits):
+ kept, used = [], -1
+ for entry in entries:
+ used += len(entry) + 1
+ if used > limit:
+ break
+ kept.append(entry)
+ fitted.append("\n".join(kept))
+ return fitted[0], fitted[1]
+
+
def _fit_decision_inputs(
question: str, plan: dict, system_chars: int, total_budget: int | None
) -> tuple[str, str]:
@@ -2347,8 +2363,7 @@ class ResearchSupervisor:
raise LeaseLost()
if resuming:
sources = list(run.get("sources") or [])[:max_sources]
- remaining = max(0, max_sources - len(sources))
- document_sources = list(run.get("documentSources") or [])[:remaining]
+ document_sources = list(run.get("documentSources") or [])[:max_sources]
for step in run.get("steps") or []:
result = step.get("result") if isinstance(step.get("result"), dict) else {}
@@ -2393,7 +2408,7 @@ class ResearchSupervisor:
or f"{source.get('documentId') or source.get('filename')}:{source.get('page') or ''}"
)
if source_key not in document_source_keys:
- if len(sources) + len(document_sources) >= max_sources:
+ if len(document_sources) >= max_sources:
continue
written = await asyncio.to_thread(
db.upsert_document_source,
@@ -2666,7 +2681,7 @@ class ResearchSupervisor:
or f"{source.get('documentId') or source.get('filename')}:{source.get('page') or ''}"
)
if source_key not in document_source_keys:
- if len(sources) + len(document_sources) >= max_sources:
+ if len(document_sources) >= max_sources:
continue
written = await asyncio.to_thread(
db.upsert_document_source,
@@ -2691,7 +2706,7 @@ class ResearchSupervisor:
rag_result = ""
rag_sources = accepted_rag_sources
for source in mcp_sources:
- if len(sources) + len(document_sources) >= max_sources:
+ if len(document_sources) >= max_sources:
break
written = await asyncio.to_thread(
db.upsert_document_source,
@@ -2705,7 +2720,7 @@ class ResearchSupervisor:
rag_sources.append(source)
step_sources = []
for match in _URL_BLOCK.finditer(result if action["action"] == "search" else ""):
- if len(sources) + len(document_sources) >= max_sources:
+ if len(sources) >= max_sources:
break
source = {k: match.group(k).strip() for k in ("title", "url", "snippet")}
allowed, _reason, _hostname = check_url_access(
@@ -2818,18 +2833,18 @@ class ResearchSupervisor:
await self._check_active(run["id"])
if not completed_steps and not sources and not document_sources:
raise ValueError(f"No research step gathered any evidence. {step_error}".rstrip())
- source_catalog = "\n".join(
+ web_entries = [
f"{index}. Title: {_citation_title(source, source['url'])}\n URL: {source['url']}"
for index, source in enumerate(sources, 1)
- )
- document_source_catalog = "\n".join(
+ ]
+ document_entries = [
f"{index}. Filename: {source.get('filename') or 'Document'}\n"
f" Page: {source.get('page') if source.get('page') is not None else '(unknown)'}\n"
f" Citation: {_document_source_citation(source)}\n"
f" Document ID: {source.get('documentId') or '(unknown)'}\n"
f" Chunk ID: {source.get('chunkId') or '(unknown)'}"
for index, source in enumerate(document_sources, 1)
- )
+ ]
# Model-derived JSON shares the evidence budget; history gets only what the fixed scaffold leaves.
total_budget = _prompt_char_budget(
_SYNTHESIS_CONTEXT_RESERVE_TOKENS, _run_inference_request(run)
@@ -2839,6 +2854,22 @@ class ResearchSupervisor:
_SYNTHESIS_AUDIT_SYSTEM_PROMPT,
run["config"],
)
+ report_system = _system_prompt_with_instructions(_REPORT_SYSTEM_PROMPT, run["config"])
+ catalog_chars = len("\n".join(web_entries)) + len("\n".join(document_entries))
+ source_catalog, document_source_catalog = _fit_synthesis_catalogs(
+ web_entries,
+ document_entries,
+ catalog_chars
+ if total_budget is None or not document_entries
+ else max(
+ 0,
+ total_budget
+ - max(len(audit_system), len(report_system))
+ - len(question)
+ - len(plan_json),
+ )
+ // 2,
+ )
audit_scaffold_chars = (
len(audit_system)
+ len(question)
@@ -2916,7 +2947,6 @@ class ResearchSupervisor:
break
except (ValueError, json.JSONDecodeError):
continue
- report_system = _system_prompt_with_instructions(_REPORT_SYSTEM_PROMPT, run["config"])
report_scaffold_chars = (
len(report_system)
+ len(question)
diff --git a/studio/backend/tests/test_research_document_evidence.py b/studio/backend/tests/test_research_document_evidence.py
index 3a857c0110c..1405789e18b 100644
--- a/studio/backend/tests/test_research_document_evidence.py
+++ b/studio/backend/tests/test_research_document_evidence.py
@@ -74,6 +74,7 @@ def _synthesis_prompt(
knowledge_base,
context_length = None,
query = "pump",
+ max_sources = 100,
) -> str:
from core import research_runs as worker
@@ -92,7 +93,7 @@ def _synthesis_prompt(
"question": "how does the pump work?",
"budgets": {
"maxSteps": STEPS,
- "maxSources": 100,
+ "maxSources": max_sources,
"modelTimeoutSeconds": 900,
"toolTimeoutSeconds": 10,
},
@@ -138,6 +139,39 @@ def test_every_step_keeps_its_top_document_passage_in_the_report_prompt(research
assert all(f"web-{step}-0" in prompt for step in range(STEPS))
+def test_web_results_do_not_use_up_the_room_for_document_passages(research_home, monkeypatch):
+ prompt = _synthesis_prompt(monkeypatch, _knowledge_base_results, max_sources = 40)
+
+ assert all(f"manual-fact-{step}" in prompt for step in range(STEPS))
+ assert all(f"Chunk ID: chunk-{step}\n" in prompt for step in range(STEPS))
+ assert len(research_db.get_run("run-1")["sources"]) == 40
+
+
+def test_small_context_keeps_passages_when_both_catalogs_fill(research_home, monkeypatch):
+ def distinct_chunks(step):
+ sources = [
+ {
+ "chunkId": f"{index:08d}-1111-4222-8333-444455556666",
+ "documentId": "9b2f6c1e-7d4a-4f0b-9c3e-2a1b5d6e7f80",
+ "filename": "manual.pdf",
+ "page": index + 1,
+ "text": _passage(index),
+ }
+ for index in range(step * 5, step * 5 + 5)
+ ]
+ return (
+ "\n\n".join(source["text"] for source in sources)
+ + RAG_SOURCES_SENTINEL
+ + json.dumps(sources)
+ )
+
+ prompt = _synthesis_prompt(monkeypatch, distinct_chunks, context_length = 8_192, max_sources = 40)
+
+ assert all(f"web-{step}-0" in prompt for step in range(STEPS))
+ assert "https://example.test/0/0" in prompt
+ assert "Chunk ID: 00000000-1111-4222-8333-444455556666" in prompt
+
+
def test_web_results_keep_their_room_when_the_documents_match_nothing(research_home, monkeypatch):
prompt = _synthesis_prompt(
monkeypatch, lambda _step: "No matching chunks were found in the knowledge base."
@@ -157,3 +191,14 @@ def test_small_context_keeps_web_results_when_the_documents_match_nothing(
)
assert all(f"web-{step}-0" in prompt for step in range(STEPS))
+
+
+def test_web_only_catalog_is_not_trimmed(research_home, monkeypatch):
+ prompt = _synthesis_prompt(
+ monkeypatch,
+ lambda _step: "No matching chunks were found in the knowledge base.",
+ context_length = 4_096,
+ )
+
+ catalog = prompt.split("<source_catalog>\n", 1)[1].split("</source_catalog>", 1)[0]
+ assert catalog.count("URL: ") == len(research_db.get_run("run-1")["sources"])