display relevant evidence

README.md

@@ -1,8 +1,5 @@
 # ClaimCheck.AI — Agentic Fact Verification for Calls
 
-<img width="1317" height="780" alt="image" src="https://github.com/user-attachments/assets/9fa259e1-3527-4573-86ec-a27053370a03" />
-
-
 **ClaimCheck.AI** is an multi-agent AI platform that turns meeting audio (Zoom/phone) into an evidence-backed report:
 1) **ASR Agent** → transcript + timestamps  
 2) **Claim Extraction (watsonx.ai LLM)** → JSON claims  
@@ -75,6 +72,7 @@ IBM_EMBEDDINGS_MODEL_ID=ibm/granite-embedding-107m-multilingual   # 384-dim
 IBM_RERANK_MODEL_ID=ibm/slate-30m-english-rtrvr-v2                 # optional
 IBM_VERIFIER_MODEL_ID=ibm/granite-3-8b-instruct
 IBM_SUMMARY_MODEL_ID=ibm/granite-3-8b-instruct
+IBM_CLAIM_MODEL_ID=ibm/granite-3-8b-instruct
 
 # Speech to Text (IBM)
 IBM_STT_URL=<your-ibm-stt-instance-url>   # e.g. https://api.us-south.speech-to-text.watson.cloud.ibm.com/instances/XXXX
@@ -101,15 +99,7 @@ If you change the KB, rebuild the index by deleting the `kb/index/` folder.
 KMP_DUPLICATE_LIB_OK=TRUE OMP_NUM_THREADS=1 uvicorn app.main:app --reload
 ```
 
-### 5) Run the UI
-```bash
-cd ui
-pnpm install
-pnpm dev
-```
-- If you change the UI dev port/host, add it to `allow_origins` in `app/main.py`.
-
-### 6) Try it
+### 5) Try it
 
 **Transcript path (no audio):**
 ```bash

app/agents/claims.py

@@ -86,7 +86,7 @@ def _build_claims_payload(transcript: str) -> Dict[str, Any]:
         "input": prompt,
         "parameters": {
             "decoding_method": "greedy",
-            "max_new_tokens": 1000,
+            "max_new_tokens": 700,
             "min_new_tokens": 0,
             "temperature": 0.0,
             "repetition_penalty": 1.0,

app/agents/retriever.py

@@ -1,10 +1,17 @@
 # app/agents/retriever.py
 from typing import List, Tuple, Dict
-import os, json, numpy as np, faiss, requests
+import os, json, re, numpy as np, faiss, requests
 
 from app.schemas.claim import Claim
 from app.schemas.evidence import Evidence
-from app.core.config import WATSONX_BASE_URL as BASE, WATSONX_PROJECT as PROJECT_ID, WATSONX_API_KEY as APIKEY, IBM_EMBEDDINGS_MODEL_ID as EMB_MODEL_ID, IBM_RERANK_MODEL_ID as RERANK_MODEL_ID, IBM_API_VERSION as VERSION, IBM_EMBEDDINGS_MODEL_ID as EMB_MODEL_ID, IBM_RERANK_MODEL_ID as RERANK_MODEL_ID, WATSONX_API_KEY as API_KEY
+from app.core.config import (
+    WATSONX_BASE_URL as BASE,
+    WATSONX_PROJECT as PROJECT_ID,
+    WATSONX_API_KEY as API_KEY,
+    IBM_EMBEDDINGS_MODEL_ID as EMB_MODEL_ID,
+    IBM_RERANK_MODEL_ID as RERANK_MODEL_ID,
+    IBM_API_VERSION as VERSION,
+)
 from app.core.auth import get_ibm_iam_token
 
 BASE_URL = (BASE or "").rstrip("/")
@@ -57,12 +64,12 @@ def _ibm_rerank(query: str, docs: list[dict], top_n: int = 5) -> list[dict]:
     url = f"{BASE_URL}/ml/v1/text/rerank?version={VERSION}"
     hdr = {"Authorization": f"Bearer {get_ibm_iam_token()}",
            "Accept":"application/json","Content-Type":"application/json"}
+    # Use stable, unique ids per passage for rerank, then map back
+    passages = [{"id": str(i), "text": d["snippet"]} for i, d in enumerate(docs)]
+    id2doc = {str(i): d for i, d in enumerate(docs)}
     payload = {
-        "input": {
-            "query": query,
-            "passages": [{"id": d["doc_id"], "text": d["snippet"]} for d in docs]
-        },
-        "model_id": RERANK_MODEL_ID,               # <-- required
+        "input": {"query": query, "passages": passages},
+        "model_id": RERANK_MODEL_ID,
         "project_id": PROJECT_ID,
         "top_n": min(top_n, len(docs))
     }
@@ -70,10 +77,9 @@ def _ibm_rerank(query: str, docs: list[dict], top_n: int = 5) -> list[dict]:
     if r.status_code != 200:
         return docs
     order = r.json().get("results", [])
-    id2doc = {d["doc_id"]: d for d in docs}
     out = []
     for it in order:
-        d = id2doc.get(it["id"])
+        d = id2doc.get(it.get("id"))
         if d:
             d = {**d, "score": it.get("relevance", d.get("score", d.get("score", 0.0)))}
             out.append(d)
@@ -126,6 +132,9 @@ def _build_or_load():
     json.dump(docs, open(META_PATH, "w"))
     return index, docs
 
+def _normalize_snippet(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").strip()).lower()
+
 def _search(query_text: str, k: int = 8) -> list[dict]:
     index, meta = _build_or_load()
     try:
@@ -146,7 +155,16 @@ def _search(query_text: str, k: int = 8) -> list[dict]:
         hits = _ibm_rerank(query_text, hits, top_n=5) if _use_ibm() else hits
     except Exception as e:
         print(f"[retriever] IBM rerank failed, using original hits: {e}")
-    return hits
+    # Deduplicate by normalized snippet text while preserving order
+    seen_snippets = set()
+    deduped = []
+    for h in hits:
+        key = _normalize_snippet(h["snippet"])
+        if key in seen_snippets:
+            continue
+        seen_snippets.add(key)
+        deduped.append(h)
+    return deduped
 
 def retrieve_evidence_for_claims(claims: List[Claim], k: int = 8) -> Tuple[List[Claim], Dict[str, List[Evidence]]]:
     claim_to_evidence: Dict[str, List[Evidence]] = {}

app/agents/summarizer.py

@@ -97,6 +97,8 @@ def make_report(
     claims: List[Claim],
     evidence_flat: List[Evidence],
     verdicts: List[Verdict],
+    *,
+    evidence_by_claim: Dict[str, List[Evidence]] | None = None,
 ) -> CallReport:
     """
     Build a CallReport:
@@ -123,7 +125,8 @@ def make_report(
             action_items=["Review claims vs. evidence and confirm metrics in source-of-truth."],
             claims=claims,
             verdicts=verdicts,
-            evidence=evidence_flat
+            evidence=evidence_flat,
+            evidence_by_claim=evidence_by_claim or {},
         )
 
     # 3) Prepare compact context + stats
@@ -198,5 +201,6 @@ def make_report(
         action_items=action_items,
         claims=claims,
         verdicts=verdicts,
-        evidence=evidence_flat
+        evidence=evidence_flat,
+        evidence_by_claim=evidence_by_claim or {},
     )

app/agents/verifier.py

@@ -36,144 +36,150 @@ Evidence catalog (doc_id -> snippet) as JSON:
 Output JSON:
 """
 
+
 def _gen(url: str, body: dict, timeout: int = 120) -> str:
-    """Low-level call to watsonx text/generation; returns raw model text."""
-    tok = get_ibm_iam_token()
-    headers = {
-        "Authorization": f"Bearer {tok}",
-        "Accept": "application/json",
-        "Content-Type": "application/json",
-    }
-    r = requests.post(url, headers=headers, json=body, timeout=timeout)
-    r.raise_for_status()
-    j = r.json()
-    res = j.get("results") or []
-    return (res[0].get("generated_text") if res else "") or ""
+	"""Low-level call to watsonx text/generation; returns raw model text."""
+	tok = get_ibm_iam_token()
+	headers = {
+		"Authorization": f"Bearer {tok}",
+		"Accept": "application/json",
+		"Content-Type": "application/json",
+	}
+	r = requests.post(url, headers=headers, json=body, timeout=timeout)
+	r.raise_for_status()
+	j = r.json()
+	res = j.get("results") or []
+	return (res[0].get("generated_text") if res else "") or ""
+
 
 def _post_generation(prompt: str) -> dict:
-    """Call model → parse with parse_json_anywhere(root='verdicts') → repair once if needed."""
-    url = f"{WATSONX_BASE_URL.rstrip('/')}/ml/v1/text/generation?version={IBM_API_VERSION}"
-    body = {
-        "input": prompt,
-        "model_id": IBM_VERIFIER_MODEL_ID,
-        "project_id": WATSONX_PROJECT,
-        "parameters": {
-            "decoding_method": "greedy",
-            "max_new_tokens": 600,
-            "min_new_tokens": 0,
-            "repetition_penalty": 1.0,
-            "temperature": 0.0,
-        },
-    }
-
-    text = _gen(url, body)
-    parsed = parse_json_anywhere(text, root_key="verdicts")
-    if parsed and parsed.get("verdicts"):
-        return parsed
-
-    # One-shot repair: coerce to strict JSON with 'verdicts' root
-    repair_body = {
-        "input": (
-            "Return ONLY valid JSON object with root key 'verdicts' "
-            "(no prose, no markdown). If invalid, fix and output JSON:\n\n" + text
-        ),
-        "model_id": IBM_VERIFIER_MODEL_ID,
-        "project_id": WATSONX_PROJECT,
-        "parameters": {
-            "decoding_method": "greedy",
-            "max_new_tokens": 400,
-            "temperature": 0.0,
-        },
-    }
-    repaired = _gen(url, repair_body)
-    reparsed = parse_json_anywhere(repaired, root_key="verdicts")
-    if reparsed and reparsed.get("verdicts"):
-        return reparsed
-
-    # Debug preview if still not parsable
-    print("[verifier] RAW OUTPUT >>>", (text or repaired)[:1000])
-    return {"verdicts": []}
+	"""Call model → parse with parse_json_anywhere(root='verdicts') → repair once if needed."""
+	url = f"{WATSONX_BASE_URL.rstrip('/')}/ml/v1/text/generation?version={IBM_API_VERSION}"
+	body = {
+		"input": prompt,
+		"model_id": IBM_VERIFIER_MODEL_ID,
+		"project_id": WATSONX_PROJECT,
+		"parameters": {
+			"decoding_method": "greedy",
+			"max_new_tokens": 600,
+			"min_new_tokens": 0,
+			"repetition_penalty": 1.0,
+			"temperature": 0.0,
+		},
+	}
+
+	text = _gen(url, body)
+	parsed = parse_json_anywhere(text, root_key="verdicts")
+	if parsed and parsed.get("verdicts"):
+		return parsed
+
+	# One-shot repair: coerce to strict JSON with 'verdicts' root
+	repair_body = {
+		"input": (
+			"Return ONLY valid JSON object with root key 'verdicts' "
+			"(no prose, no markdown). If invalid, fix and output JSON:\n\n" + text
+		),
+		"model_id": IBM_VERIFIER_MODEL_ID,
+		"project_id": WATSONX_PROJECT,
+		"parameters": {
+			"decoding_method": "greedy",
+			"max_new_tokens": 400,
+			"temperature": 0.0,
+		},
+	}
+	repaired = _gen(url, repair_body)
+	reparsed = parse_json_anywhere(repaired, root_key="verdicts")
+	if reparsed and reparsed.get("verdicts"):
+		return reparsed
+
+	# Debug preview if still not parsable
+	print("[verifier] RAW OUTPUT >>>", (text or repaired)[:1000])
+	return {"verdicts": []}
+
 
 def verify(claims: List[Claim], evidence_map: Dict[str, List[Evidence]]) -> List[Verdict]:
-    """
-    claims: list of Claim (must have .id and .text)
-    evidence_map: claim_id -> List[Evidence] (must have .doc_id, .snippet)
-    returns: List[Verdict]
-    """
-    # 1) Flatten evidence to a doc_id -> snippet catalog
-    doc_catalog: Dict[str, str] = {}
-    for lst in evidence_map.values():
-        for e in lst:
-            doc_catalog.setdefault(e.doc_id, e.snippet)
-
-    # 2) Minimal claims JSON for the LLM
-    claims_json = [{"id": c.id, "text": c.text} for c in claims]
-
-    # 3) Build prompt
-    prompt = (
-        PROMPT
-        .replace("{CLAIMS_JSON}", json.dumps(claims_json, ensure_ascii=False))
-        .replace("{EVIDENCE_JSON}", json.dumps(doc_catalog, ensure_ascii=False))
-    )
-
-    # 4) Call model + robust parse
-    try:
-        parsed = _post_generation(prompt)
-    except Exception as e:
-        # Fail-safe: mark all as insufficient
-        print(f"[verifier] generation failed: {e}")
-        return [
-            Verdict(
-                claim_id=c.id,
-                label="insufficient",
-                confidence=0.4,
-                best_evidence_id="",
-                rationale="Verifier offline; defaulting to insufficient."
-            )
-            for c in claims
-        ]
-
-    # 5) Convert to Verdict[]
-    allowed = {"supported", "refuted", "insufficient"}
-    items = parsed.get("verdicts", []) or []
-
-    # Tiebreaker: top retrieved evidence per claim
-    top_ev: Dict[str, str] = {}
-    for c in claims:
-        evs = evidence_map.get(c.id, [])
-        best = max(evs, key=lambda e: e.score, default=None)
-        top_ev[c.id] = best.doc_id if best else ""
-
-    out: List[Verdict] = []
-    for it in items:
-        cid = it.get("claim_id", "")
-        label = (it.get("label") or "").lower()
-        conf = float(it.get("confidence", 0.5))
-        cites = it.get("citation_ids") or []
-        rationale = (it.get("rationale") or "")[:300]
-        if label not in allowed:
-            label = "insufficient"
-
-        # choose best_evidence_id from cited doc_ids or fallback to top_ev
-        best_id = next((d for d in cites if d in doc_catalog), "") or top_ev.get(cid, "")
-
-        out.append(Verdict(
-            claim_id=cid,
-            label=label,
-            confidence=conf,
-            best_evidence_id=best_id,
-            rationale=rationale
-        ))
-
-    # Ensure every claim has a verdict
-    have = {v.claim_id for v in out}
-    for c in claims:
-        if c.id not in have:
-            out.append(Verdict(
-                claim_id=c.id,
-                label="insufficient",
-                confidence=0.4,
-                best_evidence_id=top_ev.get(c.id, ""),
-                rationale="No explicit verdict returned; marking as insufficient."
-            ))
-    return out
+	"""
+	claims: list of Claim (must have .id and .text)
+	evidence_map: claim_id -> List[Evidence] (must have .doc_id, .snippet)
+	returns: List[Verdict]
+	"""
+	# 1) Flatten evidence to a doc_id -> snippet catalog
+	doc_catalog: Dict[str, str] = {}
+	for lst in evidence_map.values():
+		for e in lst:
+			doc_catalog.setdefault(e.doc_id, e.snippet)
+
+	# 2) Minimal claims JSON for the LLM
+	claims_json = [{"id": c.id, "text": c.text} for c in claims]
+
+	# 3) Build prompt
+	prompt = (
+		PROMPT
+		.replace("{CLAIMS_JSON}", json.dumps(claims_json, ensure_ascii=False))
+		.replace("{EVIDENCE_JSON}", json.dumps(doc_catalog, ensure_ascii=False))
+	)
+
+	# 4) Call model + robust parse
+	try:
+		parsed = _post_generation(prompt)
+	except Exception as e:
+		# Fail-safe: mark all as insufficient
+		print(f"[verifier] generation failed: {e}")
+		return [
+			Verdict(
+				claim_id=c.id,
+				label="insufficient",
+				confidence=0.4,
+				best_evidence_id="",
+				rationale="Verifier offline; defaulting to insufficient.",
+				citation_ids=[],
+			)
+			for c in claims
+		]
+
+	# 5) Convert to Verdict[]
+	allowed = {"supported", "refuted", "insufficient"}
+	items = parsed.get("verdicts", []) or []
+
+	# Tiebreaker: top retrieved evidence per claim
+	top_ev: Dict[str, str] = {}
+	for c in claims:
+		evs = evidence_map.get(c.id, [])
+		best = max(evs, key=lambda e: e.score, default=None)
+		top_ev[c.id] = best.doc_id if best else ""
+
+	out: List[Verdict] = []
+	for it in items:
+		cid = it.get("claim_id", "")
+		label = (it.get("label") or "").lower()
+		conf = float(it.get("confidence", 0.5))
+		cites = it.get("citation_ids") or []
+		rationale = (it.get("rationale") or "")[:300]
+		if label not in allowed:
+			label = "insufficient"
+
+		# choose best_evidence_id from cited doc_ids or fallback to top_ev
+		best_id = next((d for d in cites if d in doc_catalog), "") or top_ev.get(cid, "")
+
+		out.append(Verdict(
+			claim_id=cid,
+			label=label,
+			confidence=conf,
+			best_evidence_id=best_id,
+			rationale=rationale,
+			citation_ids=cites,
+		))
+
+	# Ensure every claim has a verdict
+	have = {v.claim_id for v in out}
+	for c in claims:
+		if c.id not in have:
+			out.append(Verdict(
+				claim_id=c.id,
+				label="insufficient",
+				confidence=0.4,
+				best_evidence_id=top_ev.get(c.id, ""),
+				rationale="No explicit verdict returned; marking as insufficient.",
+				citation_ids=[],
+			))
+	return out

app/core/orchestrator.py

@@ -9,35 +9,69 @@ from app.agents.claims import extract_claims
 from app.agents.retriever import retrieve_evidence_for_claims
 from app.agents.verifier import verify
 from app.agents.summarizer import make_report
-import os
+import os, re, json
 
 os.environ.setdefault("KMP_DUPLICATE_LIB_OK", "TRUE")
 os.environ.setdefault("OMP_NUM_THREADS", "1")
 
 
+def _norm_snippet(s: str) -> str:
+	return re.sub(r"\s+", " ", (s or "").strip()).lower()
+
+
 def process_call(audio_path: Optional[str] = None, transcript: Optional[str] = None) -> CallReport:
-    print("[orchestrator] START")
-
-    segments = transcribe(audio_path) if audio_path else [
-        {"start":0.0,"end":0.0,"speaker":"A","text": transcript or ""}]
-
-    # 2) Claim extraction (IBM)
-    claims: List[Claim] = extract_claims(segments)
-    print(f"[orchestrator] Claims extracted: {len(claims)}")
-    if not claims:
-        print("[orchestrator] No claims found; building minimal report.")
-        return make_report(segments, [], [], [])
-
-    # 3) Evidence retrieval (IBM embeddings + optional rerank)
-    claims, evmap = retrieve_evidence_for_claims(claims, k=8)
-    ev_count = sum(len(v) for v in evmap.values())
-    evidence_flat = [e for lst in evmap.values() for e in lst]
-
-    verdicts: List[Verdict] = verify(claims, evmap)
-    print(f"[orchestrator] Verifier produced {len(verdicts)} verdicts")
-
-    # 5) Summarize
-    report = make_report(segments, claims, evidence_flat, verdicts)
-    print(report.call_summary)
-    print("[orchestrator] DONE")
-    return report
+	print("[orchestrator] START")
+
+	segments = transcribe(audio_path) if audio_path else [
+		{"start":0.0,"end":0.0,"speaker":"A","text": transcript or ""}]
+
+	# 2) Claim extraction (IBM)
+	claims: List[Claim] = extract_claims(segments)
+	print(f"[orchestrator] Claims extracted: {len(claims)}")
+	if not claims:
+		print("[orchestrator] No claims found; building minimal report.")
+		return make_report(segments, [], [], [], evidence_by_claim={})
+
+	# 3) Evidence retrieval (IBM embeddings + optional rerank)
+	claims, evmap = retrieve_evidence_for_claims(claims, k=8)
+	ev_count = sum(len(v) for v in evmap.values())
+
+	# Print evidence per claim (detailed)
+	print("[orchestrator] Evidence per claim (top k):")
+	for c in claims:
+		evs = evmap.get(c.id, [])
+		print(f"  - Claim {c.id}: {c.text}")
+		for i, e in enumerate(evs, 1):
+			snippet_preview = _norm_snippet(e.snippet)[:200]
+			meta_preview = ""
+			try:
+				meta_preview = json.dumps(e.metadata, ensure_ascii=False)[:160]
+			except Exception:
+				meta_preview = str(e.metadata)[:160]
+			print(f"      {i:02d}. {e.source or e.doc_id}  id={e.doc_id}  score={e.score:.2f}")
+			print(f"          {snippet_preview}")
+			if e.metadata:
+				print(f"          meta: {meta_preview}")
+
+	# Flatten and deduplicate global evidence by normalized snippet, keeping highest score
+	flat: List[Evidence] = [e for lst in evmap.values() for e in lst]
+	by_snippet: Dict[str, Evidence] = {}
+	for e in flat:
+		key = _norm_snippet(e.snippet)
+		best = by_snippet.get(key)
+		if not best or (e.score or 0.0) > (best.score or 0.0):
+			by_snippet[key] = e
+	evidence_flat = list(by_snippet.values())
+
+	verdicts: List[Verdict] = verify(claims, evmap)
+	print(f"[orchestrator] Verifier produced {len(verdicts)} verdicts")
+	print("[orchestrator] Verdicts with citations:")
+	for v in verdicts:
+		cites = getattr(v, "citation_ids", [])
+		print(f"  - {v.claim_id}: {v.label}  conf={v.confidence:.2f}  best={v.best_evidence_id}  cites={cites}")
+
+	# 5) Summarize
+	report = make_report(segments, claims, evidence_flat, verdicts, evidence_by_claim=evmap)
+	print(report.call_summary)
+	print("[orchestrator] DONE")
+	return report

app/schemas/evidence.py

@@ -1,8 +1,8 @@
 from pydantic import BaseModel
-from typing import Dict
+from typing import Dict, Any
 class Evidence(BaseModel):
     doc_id: str
     source: str
     snippet: str
     score: float
-    metadata: Dict[str, str] = {}
+    metadata: Dict[str, Any] = {}

app/schemas/report.py

@@ -1,13 +1,14 @@
 from pydantic import BaseModel
-from typing import List
+from typing import List, Dict
 from .claim import Claim
 from .evidence import Evidence
 from .verdict import Verdict
 
 class CallReport(BaseModel):
-    call_summary: str
-    claim_table: List[dict]
-    action_items: List[str] = []
-    claims: List[Claim]
-    verdicts: List[Verdict]
-    evidence: List[Evidence]
+	call_summary: str
+	claim_table: List[dict]
+	action_items: List[str] = []
+	claims: List[Claim]
+	verdicts: List[Verdict]
+	evidence: List[Evidence]
+	evidence_by_claim: Dict[str, List[Evidence]] = {}

app/schemas/verdict.py

@@ -1,7 +1,9 @@
 from pydantic import BaseModel
+from typing import List
 class Verdict(BaseModel):
-    claim_id: str
-    label: str           # supported | refuted | insufficient
-    confidence: float
-    best_evidence_id: str
-    rationale: str
+	claim_id: str
+	label: str           # supported | refuted | insufficient
+	confidence: float
+	best_evidence_id: str
+	rationale: str
+	citation_ids: List[str] = []

requirements.txt

@@ -1,5 +1,223 @@
-fastapi
-uvicorn[standard]
-pydantic>=2
-python-dotenv
-requests
+annotated-types==0.7.0
+anyio==4.10.0
+av==15.0.0
+certifi==2025.8.3
+charset-normalizer==3.4.3
+click==8.2.1
+coloredlogs==15.0.1
+ctranslate2==4.6.0
+distro==1.9.0
+faiss-cpu==1.12.0
+fastapi==0.116.1
+faster-whisper==1.2.0
+filelock==3.18.0
+flatbuffers==25.2.10
+fsspec==2025.7.0
+h11==0.16.0
+hf-xet==1.1.7
+httpcore==1.0.9
+httptools==0.6.4
+httpx==0.28.1
+huggingface-hub==0.34.4
+humanfriendly==10.0
+idna==3.10
+Jinja2==3.1.6
+jiter==0.10.0
+joblib==1.5.1
+MarkupSafe==3.0.2
+mpmath==1.3.0
+networkx==3.5
+numpy==2.3.2
+onnxruntime==1.22.1
+openai==1.99.9
+packaging==25.0
+pillow==11.3.0
+protobuf==6.31.1
+pydantic==2.11.7
+pydantic_core==2.33.2
+pydub==0.25.1
+pyobjc==11.1
+pyobjc-core==11.1
+pyobjc-framework-Accessibility==11.1
+pyobjc-framework-Accounts==11.1
+pyobjc-framework-AddressBook==11.1
+pyobjc-framework-AdServices==11.1
+pyobjc-framework-AdSupport==11.1
+pyobjc-framework-AppleScriptKit==11.1
+pyobjc-framework-AppleScriptObjC==11.1
+pyobjc-framework-ApplicationServices==11.1
+pyobjc-framework-AppTrackingTransparency==11.1
+pyobjc-framework-AudioVideoBridging==11.1
+pyobjc-framework-AuthenticationServices==11.1
+pyobjc-framework-AutomaticAssessmentConfiguration==11.1
+pyobjc-framework-Automator==11.1
+pyobjc-framework-AVFoundation==11.1
+pyobjc-framework-AVKit==11.1
+pyobjc-framework-AVRouting==11.1
+pyobjc-framework-BackgroundAssets==11.1
+pyobjc-framework-BrowserEngineKit==11.1
+pyobjc-framework-BusinessChat==11.1
+pyobjc-framework-CalendarStore==11.1
+pyobjc-framework-CallKit==11.1
+pyobjc-framework-Carbon==11.1
+pyobjc-framework-CFNetwork==11.1
+pyobjc-framework-Cinematic==11.1
+pyobjc-framework-ClassKit==11.1
+pyobjc-framework-CloudKit==11.1
+pyobjc-framework-Cocoa==11.1
+pyobjc-framework-Collaboration==11.1
+pyobjc-framework-ColorSync==11.1
+pyobjc-framework-Contacts==11.1
+pyobjc-framework-ContactsUI==11.1
+pyobjc-framework-CoreAudio==11.1
+pyobjc-framework-CoreAudioKit==11.1
+pyobjc-framework-CoreBluetooth==11.1
+pyobjc-framework-CoreData==11.1
+pyobjc-framework-CoreHaptics==11.1
+pyobjc-framework-CoreLocation==11.1
+pyobjc-framework-CoreMedia==11.1
+pyobjc-framework-CoreMediaIO==11.1
+pyobjc-framework-CoreMIDI==11.1
+pyobjc-framework-CoreML==11.1
+pyobjc-framework-CoreMotion==11.1
+pyobjc-framework-CoreServices==11.1
+pyobjc-framework-CoreSpotlight==11.1
+pyobjc-framework-CoreText==11.1
+pyobjc-framework-CoreWLAN==11.1
+pyobjc-framework-CryptoTokenKit==11.1
+pyobjc-framework-DataDetection==11.1
+pyobjc-framework-DeviceCheck==11.1
+pyobjc-framework-DeviceDiscoveryExtension==11.1
+pyobjc-framework-DictionaryServices==11.1
+pyobjc-framework-DiscRecording==11.1
+pyobjc-framework-DiscRecordingUI==11.1
+pyobjc-framework-DiskArbitration==11.1
+pyobjc-framework-DVDPlayback==11.1
+pyobjc-framework-EventKit==11.1
+pyobjc-framework-ExceptionHandling==11.1
+pyobjc-framework-ExecutionPolicy==11.1
+pyobjc-framework-ExtensionKit==11.1
+pyobjc-framework-ExternalAccessory==11.1
+pyobjc-framework-FileProvider==11.1
+pyobjc-framework-FileProviderUI==11.1
+pyobjc-framework-FinderSync==11.1
+pyobjc-framework-FSEvents==11.1
+pyobjc-framework-FSKit==11.1
+pyobjc-framework-GameCenter==11.1
+pyobjc-framework-GameController==11.1
+pyobjc-framework-GameKit==11.1
+pyobjc-framework-GameplayKit==11.1
+pyobjc-framework-HealthKit==11.1
+pyobjc-framework-ImageCaptureCore==11.1
+pyobjc-framework-InputMethodKit==11.1
+pyobjc-framework-InstallerPlugins==11.1
+pyobjc-framework-InstantMessage==11.1
+pyobjc-framework-Intents==11.1
+pyobjc-framework-IntentsUI==11.1
+pyobjc-framework-IOBluetooth==11.1
+pyobjc-framework-IOBluetoothUI==11.1
+pyobjc-framework-IOSurface==11.1
+pyobjc-framework-iTunesLibrary==11.1
+pyobjc-framework-KernelManagement==11.1
+pyobjc-framework-LatentSemanticMapping==11.1
+pyobjc-framework-LaunchServices==11.1
+pyobjc-framework-libdispatch==11.1
+pyobjc-framework-libxpc==11.1
+pyobjc-framework-LinkPresentation==11.1
+pyobjc-framework-LocalAuthentication==11.1
+pyobjc-framework-LocalAuthenticationEmbeddedUI==11.1
+pyobjc-framework-MailKit==11.1
+pyobjc-framework-MapKit==11.1
+pyobjc-framework-MediaAccessibility==11.1
+pyobjc-framework-MediaExtension==11.1
+pyobjc-framework-MediaLibrary==11.1
+pyobjc-framework-MediaPlayer==11.1
+pyobjc-framework-MediaToolbox==11.1
+pyobjc-framework-Metal==11.1
+pyobjc-framework-MetalFX==11.1
+pyobjc-framework-MetalKit==11.1
+pyobjc-framework-MetalPerformanceShaders==11.1
+pyobjc-framework-MetalPerformanceShadersGraph==11.1
+pyobjc-framework-MetricKit==11.1
+pyobjc-framework-MLCompute==11.1
+pyobjc-framework-ModelIO==11.1
+pyobjc-framework-MultipeerConnectivity==11.1
+pyobjc-framework-NaturalLanguage==11.1
+pyobjc-framework-NetFS==11.1
+pyobjc-framework-Network==11.1
+pyobjc-framework-NetworkExtension==11.1
+pyobjc-framework-NotificationCenter==11.1
+pyobjc-framework-OpenDirectory==11.1
+pyobjc-framework-OSAKit==11.1
+pyobjc-framework-OSLog==11.1
+pyobjc-framework-PassKit==11.1
+pyobjc-framework-PencilKit==11.1
+pyobjc-framework-PHASE==11.1
+pyobjc-framework-Photos==11.1
+pyobjc-framework-PhotosUI==11.1
+pyobjc-framework-PreferencePanes==11.1
+pyobjc-framework-PushKit==11.1
+pyobjc-framework-Quartz==11.1
+pyobjc-framework-QuickLookThumbnailing==11.1
+pyobjc-framework-ReplayKit==11.1
+pyobjc-framework-SafariServices==11.1
+pyobjc-framework-SafetyKit==11.1
+pyobjc-framework-SceneKit==11.1
+pyobjc-framework-ScreenCaptureKit==11.1
+pyobjc-framework-ScreenSaver==11.1
+pyobjc-framework-ScreenTime==11.1
+pyobjc-framework-ScriptingBridge==11.1
+pyobjc-framework-SearchKit==11.1
+pyobjc-framework-Security==11.1
+pyobjc-framework-SecurityFoundation==11.1
+pyobjc-framework-SecurityInterface==11.1
+pyobjc-framework-SecurityUI==11.1
+pyobjc-framework-SensitiveContentAnalysis==11.1
+pyobjc-framework-ServiceManagement==11.1
+pyobjc-framework-SharedWithYou==11.1
+pyobjc-framework-SharedWithYouCore==11.1
+pyobjc-framework-ShazamKit==11.1
+pyobjc-framework-Social==11.1
+pyobjc-framework-SoundAnalysis==11.1
+pyobjc-framework-Speech==11.1
+pyobjc-framework-SpriteKit==11.1
+pyobjc-framework-StoreKit==11.1
+pyobjc-framework-Symbols==11.1
+pyobjc-framework-SyncServices==11.1
+pyobjc-framework-SystemConfiguration==11.1
+pyobjc-framework-SystemExtensions==11.1
+pyobjc-framework-ThreadNetwork==11.1
+pyobjc-framework-UniformTypeIdentifiers==11.1
+pyobjc-framework-UserNotifications==11.1
+pyobjc-framework-UserNotificationsUI==11.1
+pyobjc-framework-VideoSubscriberAccount==11.1
+pyobjc-framework-VideoToolbox==11.1
+pyobjc-framework-Virtualization==11.1
+pyobjc-framework-Vision==11.1
+pyobjc-framework-WebKit==11.1
+python-dotenv==1.1.1
+python-multipart==0.0.20
+pyttsx3==2.99
+PyYAML==6.0.2
+regex==2025.7.34
+requests==2.32.4
+safetensors==0.6.2
+scikit-learn==1.7.1
+scipy==1.16.1
+sentence-transformers==5.1.0
+setuptools==80.9.0
+sniffio==1.3.1
+starlette==0.47.2
+sympy==1.14.0
+threadpoolctl==3.6.0
+tokenizers==0.21.4
+torch==2.8.0
+tqdm==4.67.1
+transformers==4.55.2
+typing-inspection==0.4.1
+typing_extensions==4.14.1
+urllib3==2.5.0
+uvicorn==0.35.0
+uvloop==0.21.0
+watchfiles==1.1.0
+websockets==15.0.1

ui/claimcheck-ui/src/components/ClaimsTable.tsx

@@ -25,7 +25,9 @@ export function ClaimsTable() {
     data.sort((a: any, b: any) => (sortAsc ? (a.verdict?.confidence ?? 0) - (b.verdict?.confidence ?? 0) : (b.verdict?.confidence ?? 0) - (a.verdict?.confidence ?? 0)))
 
     return data.map((r: any) => {
-      const evid = evidence.find((e) => e.doc_id === r.verdict?.best_evidence_id)
+      const cited = r.verdict?.citation_ids as string[] | undefined
+      const firstId = (cited && cited[0]) || r.verdict?.best_evidence_id
+      const evid = firstId ? evidence.find((e) => e.doc_id === firstId) : undefined
       return { ...r, evidence: evid }
     })
   }, [run, filter, sortAsc])

ui/claimcheck-ui/src/components/EvidenceDrawer.tsx

@@ -1,5 +1,6 @@
 import { useMemo } from 'react'
 import { useAppStore } from '@/store/appStore'
+import type { Evidence } from '@/lib/types'
 
 export function EvidenceDrawer() {
   const run = useAppStore((s) => s.runs.find((r) => r.id === s.currentRunId))
@@ -11,10 +12,24 @@ export function EvidenceDrawer() {
   const verdict = useMemo(() => run?.report?.verdicts.find((v) => v.claim_id === claimId), [run, claimId])
 
   const evidences = useMemo(() => {
-    if (!run?.report || !verdict) return []
-    const sorted = [...run.report.evidence].sort((a, b) => (b.score ?? 0) - (a.score ?? 0))
+    if (!run?.report) return [] as Evidence[]
+    const report = run.report
+
+    // 1) Prefer cited evidence for this claim
+    const citedIds = verdict?.citation_ids || []
+    if (citedIds.length > 0) {
+      const cited = report.evidence.filter((e) => citedIds.includes(e.doc_id))
+      if (cited.length > 0) return cited
+    }
+
+    // 2) Fallback to per-claim retrieved evidence
+    const byClaim = report.evidence_by_claim?.[claimId ?? ''] || []
+    if (byClaim.length > 0) return byClaim
+
+    // 3) Final fallback: top-scored global evidence
+    const sorted = [...report.evidence].sort((a, b) => (b.score ?? 0) - (a.score ?? 0))
     return sorted.slice(0, 5)
-  }, [run, verdict])
+  }, [run, verdict, claimId])
 
   if (!isOpen) return null
 
@@ -37,8 +52,8 @@ export function EvidenceDrawer() {
           <div>
             <div className="text-xs text-muted-foreground">Top Evidence</div>
             <ul className="mt-2 space-y-3">
-              {evidences.map((e) => (
-                <li key={e.doc_id} className="rounded-md border p-3">
+              {evidences.map((e, idx) => (
+                <li key={`${e.doc_id}-${idx}`} className="rounded-md border p-3">
                   <div className="text-xs text-muted-foreground">{e.source || e.doc_id} • score {(e.score ?? 0).toFixed(2)}</div>
                   <div className="mt-1 text-sm whitespace-pre-wrap">{e.snippet}</div>
                   {e.metadata && (

ui/claimcheck-ui/src/lib/types.ts

@@ -23,6 +23,7 @@ export type Verdict = {
   confidence?: number
   best_evidence_id?: string
   rationale?: string
+  citation_ids?: string[]
 }
 
 export type CallReport = {
@@ -32,6 +33,7 @@ export type CallReport = {
   claims: Claim[]
   evidence: Evidence[]
   verdicts: Verdict[]
+  evidence_by_claim?: Record<string, Evidence[]>
 }
 
 export type InputKind = 'audio' | 'transcript' | 'sample'