Spaces:
Running
Running
File size: 8,721 Bytes
4939f56 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 | """Final synthesis — techspec.md §3.
Builds a single structured "Final Report" for a completed pipeline run:
one headline, per-finding claims each tagged with the run's confidence tier
(identified / homolog / de_novo), and honest caveats.
An LLM pass polishes the wording when a provider is reachable; the findings
and caveats themselves are always assembled deterministically from real step
results, so the panel never invents content and never blocks the pipeline.
"""
import logging
logger = logging.getLogger(__name__)
_CAVEATS = {
"identified": [
"Annotations describe the matched database entry — transferability to your exact input depends on sequence identity.",
],
"homolog": [
"The query was resolved via sequence similarity, so annotations describe a homolog, not the exact protein.",
"Verify functional transferability before relying on homolog-derived annotations experimentally.",
],
"de_novo": [
"No database match was found — every annotation is a computational prediction on the raw sequence.",
"Treat composition- and structure-based hints as hypotheses for experimental follow-up.",
],
}
def _tier(context: dict) -> str:
return (context.get("query") or {}).get("confidence") or "identified"
def build_findings(context: dict) -> tuple[list[dict], list[dict]]:
"""Deterministic findings + stats from actual step results."""
tier = _tier(context)
findings: list[dict] = []
stats = context.get("query", {})
blast = context.get("blast") or {}
top_hit = blast.get("top_hit")
if top_hit:
findings.append({
"claim": f"Closest database match: {top_hit.get('description', 'unknown')} ({top_hit.get('accession', '?')})",
"confidence_tier": tier,
"source_tool": "blast",
"page_url": f"https://www.ncbi.nlm.nih.gov/protein/{top_hit.get('accession', '')}" if top_hit.get("accession") else None,
})
elif blast.get("count", 0) == 0:
findings.append({
"claim": "No significant similarity to any database sequence.",
"confidence_tier": tier,
"source_tool": "blast",
"page_url": None,
})
uniprot = context.get("uniprot") or {}
if uniprot.get("_de_novo"):
comp = uniprot.get("composition") or {}
findings.append({
"claim": (
f"De novo characterization: {comp.get('sequence_type', 'protein')} of "
f"{comp.get('length', stats.get('length', '?'))} residues; function hints are heuristic only."
),
"confidence_tier": tier,
"source_tool": "uniprot",
"page_url": None,
})
elif uniprot and not uniprot.get("error"):
name = uniprot.get("full_name") or uniprot.get("accession") or "entry"
findings.append({
"claim": f"Annotated as {name} ({uniprot.get('accession', '?')}) in UniProt.",
"confidence_tier": tier,
"source_tool": "uniprot",
"page_url": f"https://www.uniprot.org/uniprotkb/{uniprot.get('accession', '')}" if uniprot.get("accession") else None,
})
domains = context.get("domains") or {}
dom_list = domains.get("domains") or []
if dom_list:
top_domain = dom_list[0]
findings.append({
"claim": (
f"{len(dom_list)} domain(s) detected — top: {top_domain.get('name', '?')} "
f"({top_domain.get('source_db', '?')})."
),
"confidence_tier": tier,
"source_tool": "domains",
"page_url": (
f"https://www.ebi.ac.uk/interpro/entry/InterPro/{top_domain['accession']}"
if top_domain.get("accession") else None
),
})
msa = context.get("msa") or {}
if msa.get("aln_fasta"):
findings.append({
"claim": f"Multiple sequence alignment built over {msa.get('sequence_count', '?')} sequences.",
"confidence_tier": tier,
"source_tool": "msa",
"page_url": None,
})
pathway = context.get("pathway_enrichment") or {}
pw_list = (pathway.get("pathways") if isinstance(pathway, dict) else None) or []
if pw_list:
top_pw = pw_list[0]
st_id = top_pw.get("stId", "")
findings.append({
"claim": (
f"Pathway enrichment: {len(pw_list)} Reactome pathway(s), "
f"e.g. {top_pw.get('name', '?')} (FDR {top_pw.get('entitiesFDR', '?')})."
),
"confidence_tier": tier,
"source_tool": "pathway_enrichment",
"page_url": f"https://reactome.org/PathwayBrowser/#{st_id}" if st_id else None,
})
elif isinstance(pathway, dict) and pathway.get("error"):
findings.append({
"claim": f"Pathway enrichment unavailable ({str(pathway.get('error'))[:80]}).",
"confidence_tier": tier,
"source_tool": "pathway_enrichment",
"page_url": None,
})
af = context.get("alphafold") or {}
if af.get("structure_available"):
source_label = "ESMFold prediction" if af.get("source") == "esmfold" else "AlphaFold DB model"
plddt = af.get("mean_plddt") or af.get("confidence")
detail = f", mean pLDDT {plddt}" if plddt else ""
findings.append({
"claim": f"3D structure available ({source_label}{detail}).",
"confidence_tier": tier,
"source_tool": "alphafold",
"page_url": af.get("pdb_url"),
})
return findings, [f"Confidence tier for this run: {tier}."] + _CAVEATS.get(tier, [])
def _headline(context: dict, findings: list[dict]) -> str:
tier = _tier(context)
if tier == "de_novo":
return "Novel sequence — characterized de novo from composition and predicted structure."
top = next((f for f in findings if f["source_tool"] == "uniprot"), None)
if top:
return f"Identified with {tier}-confidence annotation: {top['claim'].removeprefix('Annotated as ').rstrip('.')}"
hit = next((f for f in findings if f["source_tool"] == "blast"), None)
return hit["claim"] if hit else "Analysis complete."
async def _polish_with_llm(headline: str, summary_parts: list[str]) -> str | None:
"""Best-effort LLM rewrite of the headline+summary. Returns None on failure."""
try:
from app.ai.interpreter import _get_acompletion
from app.ai import llm_client
acompletion = _get_acompletion()
candidates = llm_client.get_all_candidates()
if acompletion is None or not candidates:
return None
prompt = (
"Rewrite the following research summary into ONE crisp paragraph (max 90 words).\n"
"Do not add facts that are not present. Keep hedged language.\n\n"
f"Headline: {headline}\nFindings:\n- " + "\n- ".join(summary_parts)
)
candidate = candidates[0]
response = await acompletion(
model=candidate["model"],
messages=[{"role": "user", "content": prompt}],
temperature=0.3,
max_tokens=300,
timeout=20,
api_key=candidate["api_key"],
)
text = response.choices[0].message.content if response.choices else ""
return text.strip() or None
except Exception as exc:
logger.info("Synthesis LLM polish skipped: %s", exc)
return None
async def synthesize(context: dict) -> dict:
"""Produce the final_report block stored in context_json."""
findings, caveats = build_findings(context)
headline = _headline(context, findings)
polished = await _polish_with_llm(
headline, [f["claim"] for f in findings] if findings else ["No significant results."]
)
report = {
"headline": headline,
"summary": polished or " ".join(f["claim"] for f in findings if f["source_tool"] == "blast")
or "Analysis complete — see per-tool sections below.",
"findings": findings,
"caveats": caveats,
"_mode": "llm_polished" if polished else "deterministic",
}
logger.info(
"Final synthesis built (%s): %d findings",
report["_mode"],
len(findings),
)
return report
def synthesize_sync(context: dict) -> dict:
"""Synchronous variant (no LLM polish) for export paths."""
findings, caveats = build_findings(context)
return {
"headline": _headline(context, findings),
"summary": " ".join(f["claim"] for f in findings[:3]),
"findings": findings,
"caveats": caveats,
"_mode": "deterministic",
}
|