RFTSystems's picture
Create app.py
f602dc7 verified
import os
import re
import io
import json
import time
import uuid
import zipfile
import sqlite3
import hashlib
from dataclasses import dataclass
from typing import List, Dict, Any, Optional, Tuple
import gradio as gr
BASE_DIR = os.environ.get("RFT_MEM_BASE", "var/rftmem")
os.makedirs(BASE_DIR, exist_ok=True)
def sha256_str(s: str) -> str:
return hashlib.sha256(s.encode("utf-8")).hexdigest()
def now_ms() -> int:
return int(time.time() * 1000)
def atomic_write(path: str, data: bytes) -> None:
tmp = path + ".tmp"
with open(tmp, "wb") as f:
f.write(data)
f.flush()
os.fsync(f.fileno())
os.replace(tmp, path)
def safe_fts_match(user_query: str) -> str:
words = re.findall(r"[A-Za-z0-9_]+", (user_query or "").lower())
if not words:
return "___NO_HITS___"
seen = set()
uniq = []
for w in words:
if w not in seen:
seen.add(w)
uniq.append(w)
return " OR ".join(uniq)
@dataclass
class RetrievalHit:
event_id: str
seq: int
role: str
text: str
ts_ms: int
digest: str
chain_hash: str
score: float
class RFTMemoryStore:
"""
Append-only ledger + SQLite FTS retrieval + hash-chained integrity.
Produces per-turn receipts that can be verified against stored events.
"""
def __init__(self, base_dir: str):
self.base_dir = base_dir
self.db_path = os.path.join(base_dir, "index.sqlite")
self._init_db()
def _connect(self) -> sqlite3.Connection:
return sqlite3.connect(self.db_path)
def _init_db(self):
os.makedirs(self.base_dir, exist_ok=True)
con = self._connect()
cur = con.cursor()
cur.execute("""
CREATE TABLE IF NOT EXISTS events (
session_id TEXT,
event_id TEXT PRIMARY KEY,
seq INTEGER,
ts_ms INTEGER,
role TEXT,
text TEXT,
digest TEXT,
prev_hash TEXT,
chain_hash TEXT,
collapse REAL
)
""")
cur.execute("""
CREATE TABLE IF NOT EXISTS receipts (
receipt_id TEXT PRIMARY KEY,
session_id TEXT,
ts_ms INTEGER,
prompt_hash TEXT,
response_hash TEXT,
receipt_path TEXT
)
""")
# Ensure join-safe FTS5 (stored content)
cur.execute("SELECT sql FROM sqlite_master WHERE type='table' AND name='events_fts'")
row = cur.fetchone()
needs_rebuild = False
if row is None:
needs_rebuild = True
else:
sql = (row[0] or "").lower()
if "content=''" in sql or 'content=""' in sql:
needs_rebuild = True
if needs_rebuild:
cur.execute("DROP TABLE IF EXISTS events_fts")
cur.execute("""
CREATE VIRTUAL TABLE events_fts USING fts5(
event_id,
session_id,
text
)
""")
con.commit()
cur.execute("DELETE FROM events_fts")
cur.execute("""
INSERT INTO events_fts(event_id, session_id, text)
SELECT event_id, session_id, text FROM events
""")
con.commit()
con.close()
# Filesystem
def session_dir(self, session_id: str) -> str:
d = os.path.join(self.base_dir, "sessions", session_id)
os.makedirs(d, exist_ok=True)
return d
def session_log_path(self, session_id: str) -> str:
return os.path.join(self.session_dir(session_id), "events.jsonl")
def receipts_dir(self, session_id: str) -> str:
d = os.path.join(self.session_dir(session_id), "receipts")
os.makedirs(d, exist_ok=True)
return d
def exports_dir(self, session_id: str) -> str:
d = os.path.join(self.session_dir(session_id), "exports")
os.makedirs(d, exist_ok=True)
return d
# Ledger ops
def get_events(self, session_id: str, limit: int = 600) -> List[Dict[str, Any]]:
con = self._connect()
cur = con.cursor()
cur.execute("""
SELECT event_id, seq, ts_ms, role, text, digest, prev_hash, chain_hash, collapse
FROM events
WHERE session_id=?
ORDER BY seq ASC
LIMIT ?
""", (session_id, int(limit)))
rows = cur.fetchall()
con.close()
out = []
for r in rows:
out.append({
"event_id": r[0],
"seq": int(r[1] or 0),
"ts_ms": r[2],
"role": r[3],
"text": r[4],
"digest": r[5],
"prev_hash": r[6],
"chain_hash": r[7],
"collapse": float(r[8] or 0.0),
})
return out
def _get_last_seq_and_chain(self, session_id: str) -> Tuple[int, str]:
con = self._connect()
cur = con.cursor()
cur.execute("SELECT COALESCE(MAX(seq), 0) FROM events WHERE session_id=?", (session_id,))
last_seq = int(cur.fetchone()[0] or 0)
cur.execute("""
SELECT chain_hash FROM events
WHERE session_id=?
ORDER BY seq DESC
LIMIT 1
""", (session_id,))
row = cur.fetchone()
con.close()
last_chain = row[0] if row and row[0] else ("0" * 64)
return last_seq, last_chain
def collapse_score(self, session_id: str, role: str, text: str) -> float:
role_w = {"user": 1.0, "tool": 0.9, "assistant": 0.6}.get(role, 0.7)
tokens = set(t.lower() for t in re.findall(r"[A-Za-z0-9_]+", text or ""))
if not tokens:
return 0.0
recent = self.get_events(session_id, limit=20)
recent_tokens = set()
for e in recent:
recent_tokens |= set(t.lower() for t in re.findall(r"[A-Za-z0-9_]+", e["text"] or ""))
unseen = len(tokens - recent_tokens)
novelty = unseen / max(1, len(tokens))
length_factor = min(1.0, len(tokens) / 30.0)
score = role_w * (0.65 * novelty + 0.35 * length_factor)
return float(max(0.0, min(1.0, score)))
def append_event(self, session_id: str, role: str, text: str) -> Dict[str, Any]:
event_id = uuid.uuid4().hex
ts = now_ms()
last_seq, prev_chain = self._get_last_seq_and_chain(session_id)
seq = last_seq + 1
payload = {
"session_id": session_id,
"event_id": event_id,
"seq": seq,
"ts_ms": ts,
"role": role,
"text": text
}
digest = sha256_str(json.dumps(payload, sort_keys=True, ensure_ascii=False))
chain_hash = sha256_str(prev_chain + digest)
collapse = self.collapse_score(session_id, role, text)
rec = {**payload, "digest": digest, "prev_hash": prev_chain, "chain_hash": chain_hash, "collapse": collapse}
# JSONL source of truth
log_path = self.session_log_path(session_id)
line = (json.dumps(rec, ensure_ascii=False) + "\n").encode("utf-8")
with open(log_path, "ab") as f:
f.write(line)
f.flush()
os.fsync(f.fileno())
# Index
con = self._connect()
cur = con.cursor()
cur.execute("""
INSERT INTO events(session_id,event_id,seq,ts_ms,role,text,digest,prev_hash,chain_hash,collapse)
VALUES(?,?,?,?,?,?,?,?,?,?)
""", (session_id, event_id, seq, ts, role, text, digest, prev_chain, chain_hash, collapse))
cur.execute("INSERT INTO events_fts(event_id, session_id, text) VALUES(?,?,?)", (event_id, session_id, text))
con.commit()
con.close()
return rec
# Retrieval
def search_lexical(self, session_id: str, query: str, k: int = 8) -> List[RetrievalHit]:
match = safe_fts_match(query)
if match == "___NO_HITS___":
return []
con = self._connect()
cur = con.cursor()
cur.execute("""
SELECT e.event_id, e.seq, e.role, e.text, e.ts_ms, e.digest, e.chain_hash,
bm25(events_fts) as rank
FROM events_fts
JOIN events e ON e.event_id = events_fts.event_id
WHERE events_fts.text MATCH ? AND e.session_id=?
ORDER BY rank ASC
LIMIT ?
""", (match, session_id, int(k)))
rows = cur.fetchall()
con.close()
hits: List[RetrievalHit] = []
for (eid, seq, role, text, ts, digest, chain_hash, rank) in rows:
r = float(rank if rank is not None else 0.0)
score = 1.0 / (1.0 + max(0.0, r))
hits.append(RetrievalHit(eid, int(seq or 0), role, text, ts, digest, chain_hash, score))
# UX fallback (if someone searches a single token that FTS doesn't match as expected)
if not hits:
tokens = re.findall(r"[A-Za-z0-9_]+", (query or "").lower())
if tokens:
needle = f"%{tokens[-1]}%"
con = self._connect()
cur = con.cursor()
cur.execute("""
SELECT event_id, seq, role, text, ts_ms, digest, chain_hash
FROM events
WHERE session_id=? AND LOWER(text) LIKE ?
ORDER BY seq DESC
LIMIT ?
""", (session_id, needle, int(k)))
rows2 = cur.fetchall()
con.close()
for (eid, seq, role, text, ts, digest, chain_hash) in rows2:
hits.append(RetrievalHit(eid, int(seq or 0), role, text, ts, digest, chain_hash, 0.001))
return hits
# Receipts
def write_receipt(self, session_id: str, user_text: str, retrieved: List[RetrievalHit], prompt: str, response: str) -> str:
receipt_id = uuid.uuid4().hex
ts = now_ms()
receipt = {
"receipt_id": receipt_id,
"session_id": session_id,
"ts_ms": ts,
"query": user_text,
"retrieval": [{
"event_id": h.event_id,
"seq": h.seq,
"role": h.role,
"content": h.text,
"score": h.score,
"digest": h.digest,
"chain_hash": h.chain_hash
} for h in retrieved],
"prompt_hash": sha256_str(prompt),
"response_hash": sha256_str(response),
"engine": {"name": "RFTSystems TrustStack", "version": "1.0", "method": "ledger + fts + receipts + guardrails"}
}
path = os.path.join(self.receipts_dir(session_id), f"{receipt_id}.json")
atomic_write(path, json.dumps(receipt, indent=2, ensure_ascii=False).encode("utf-8"))
con = self._connect()
cur = con.cursor()
cur.execute("""
INSERT INTO receipts(receipt_id, session_id, ts_ms, prompt_hash, response_hash, receipt_path)
VALUES(?,?,?,?,?,?)
""", (receipt_id, session_id, ts, receipt["prompt_hash"], receipt["response_hash"], path))
con.commit()
con.close()
return path
def verify_receipt(self, receipt_json: Dict[str, Any]) -> Tuple[bool, str]:
session_id = receipt_json.get("session_id")
if not session_id:
return False, "Missing session_id."
con = self._connect()
cur = con.cursor()
for item in receipt_json.get("retrieval", []):
eid = item.get("event_id")
expected_digest = item.get("digest")
expected_chain = item.get("chain_hash")
cur.execute("SELECT digest, chain_hash FROM events WHERE event_id=? AND session_id=?", (eid, session_id))
row = cur.fetchone()
if not row:
con.close()
return False, f"Event not found: {eid}"
if row[0] != expected_digest:
con.close()
return False, f"Digest mismatch: {eid}"
if row[1] != expected_chain:
con.close()
return False, f"Chain hash mismatch: {eid}"
con.close()
return True, "Receipt verified: all referenced events exist and hashes match."
# Trace export (OTel-style JSON, but self-contained)
def export_trace(self, session_id: str, receipt_path: str) -> str:
export_id = uuid.uuid4().hex
ts = now_ms()
with open(receipt_path, "r", encoding="utf-8") as f:
receipt = json.load(f)
trace = {
"trace_id": sha256_str(session_id + receipt.get("receipt_id", "") + str(ts)),
"session_id": session_id,
"ts_ms": ts,
"spans": [
{"span": "user.turn", "attrs": {"query": receipt.get("query", ""), "receipt_id": receipt.get("receipt_id", "")}},
{"span": "memory.retrieve", "attrs": {"k": len(receipt.get("retrieval", [])), "retrieval": receipt.get("retrieval", [])}},
{"span": "assistant.respond", "attrs": {"prompt_hash": receipt.get("prompt_hash", ""), "response_hash": receipt.get("response_hash", "")}},
{"span": "receipt.write", "attrs": {"receipt_id": receipt.get("receipt_id", ""), "receipt_path": receipt_path}},
]
}
out_path = os.path.join(self.exports_dir(session_id), f"trace_{export_id}.json")
atomic_write(out_path, json.dumps(trace, indent=2, ensure_ascii=False).encode("utf-8"))
return out_path
# Audit pack export
def export_audit_pack(self, session_id: str) -> str:
export_id = uuid.uuid4().hex
out_zip = os.path.join(self.exports_dir(session_id), f"audit_pack_{export_id}.zip")
events = self.get_events(session_id, limit=5000)
receipts_dir = self.receipts_dir(session_id)
ledger_path = self.session_log_path(session_id)
summary = {
"session_id": session_id,
"export_id": export_id,
"ts_ms": now_ms(),
"events_count": len(events),
"last_chain_hash": (events[-1]["chain_hash"] if events else "0"*64),
"integrity": "hash-chained ledger + receipt verification",
}
with zipfile.ZipFile(out_zip, "w", compression=zipfile.ZIP_DEFLATED) as z:
# include JSONL ledger if present
if os.path.exists(ledger_path):
z.write(ledger_path, arcname="ledger/events.jsonl")
# include receipts
for fn in os.listdir(receipts_dir):
if fn.endswith(".json"):
z.write(os.path.join(receipts_dir, fn), arcname=f"receipts/{fn}")
# include summary
z.writestr("summary.json", json.dumps(summary, indent=2, ensure_ascii=False))
return out_zip
store = RFTMemoryStore(BASE_DIR)
INVESTOR_SCRIPT = [
"Store these exactly: Dog=Nova, City=Manchester, Drink=Pepsi Max.",
"What is my dog's name?",
"What city did I say?",
"My drink is Coke Zero now. This overrides earlier.",
"What is my favourite drink?",
"Search for: Nova",
]
EXAMPLE_PROMPTS = [
"Store this: Dog=Nova, City=Manchester, Drink=Pepsi Max.",
"What is my dog's name?",
"What city did I say?",
"My drink is Coke Zero now. This overrides earlier.",
"What is my favourite drink?",
"Search for: Nova",
"Search for: Manchester",
]
def new_session_id() -> str:
return uuid.uuid4().hex
def events_to_messages(events: List[Dict[str, Any]]) -> List[Dict[str, str]]:
return [{"role": e["role"], "content": e["text"]} for e in events if e["role"] in ("user", "assistant")]
def format_ledger(events: List[Dict[str, Any]]) -> str:
lines = []
for e in events[-250:]:
t = time.strftime("%Y-%m-%d %H:%M:%S", time.gmtime(e["ts_ms"] / 1000))
lines.append(
f"{t} | seq={e['seq']} | {e['role']}\n"
f"{e['text']}\n"
f"event_id={e['event_id']} collapse={e['collapse']:.2f}\n"
f"digest={e['digest']}\n"
f"chain={e['chain_hash']}\n"
f"{'-'*72}"
)
return "\n".join(lines)
def build_prompt(user_msg: str, hits: List[RetrievalHit]) -> str:
memories = "\n".join([f"- ({h.role}) {h.text}" for h in hits]) if hits else "(none)"
return (
"SYSTEM: Use retrieved memory slices if relevant. Prefer exact stored facts.\n"
f"RETRIEVED MEMORIES:\n{memories}\n\n"
f"USER:\n{user_msg}\n"
)
def extract_fact_from_hits(hits: List[RetrievalHit], key: str) -> Optional[str]:
key_l = key.lower()
patterns = [
rf"\b{re.escape(key_l)}\b\s*=\s*([A-Za-z0-9 _\-']+)",
rf"\b{re.escape(key_l)}\b\s*:\s*([A-Za-z0-9 _\-']+)",
]
for h in hits:
tl = (h.text or "").strip().lower()
if key_l == "name":
m = re.search(r"\bmy name is\b\s+([A-Za-z0-9 _\-']+)", tl)
if m:
return m.group(1).strip().title()
for p in patterns:
m = re.search(p, tl)
if m:
return m.group(1).strip().strip(",.")
return None
def answer_from_memory(user_msg: str, hits: List[RetrievalHit]) -> str:
q = (user_msg or "").lower()
if q.startswith("search for") or q.startswith("search:"):
if not hits:
return "No matching memory slices were retrieved for this search."
return "Search hits:\n" + "\n".join([f"- {h.score:.4f} | {h.role} | {h.text}" for h in hits])
if "dog" in q and "name" in q:
v = extract_fact_from_hits(hits, "dog")
return f"Your dog’s name (from stored memory) is: {v}" if v else "I didn’t retrieve a stored dog name for this query."
if "city" in q:
v = extract_fact_from_hits(hits, "city")
return f"Your city (from stored memory) is: {v}" if v else "I didn’t retrieve a stored city for this query."
if "drink" in q:
v = extract_fact_from_hits(hits, "drink")
return f"Your drink (from stored memory) is: {v}" if v else "I didn’t retrieve a stored drink for this query."
if not hits:
return "No matching memory slices were retrieved for this query."
return "Retrieved memory slices:\n" + "\n".join([f"- {h.score:.4f} | {h.role} | {h.text}" for h in hits])
def chat_turn(session_id: str, user_msg: str, retrieval_k: int):
if not session_id:
session_id = new_session_id()
store.append_event(session_id, "user", user_msg)
hits = store.search_lexical(session_id, user_msg, k=int(retrieval_k))
prompt = build_prompt(user_msg, hits)
response = answer_from_memory(user_msg, hits)
store.append_event(session_id, "assistant", response)
receipt_path = store.write_receipt(session_id, user_msg, hits, prompt, response)
events = store.get_events(session_id, limit=1200)
ledger = format_ledger(events)
retrieved_view = "\n".join([f"{h.score:.4f} | {h.role} | {h.text}" for h in hits]) if hits else "(none)"
messages = events_to_messages(events)
# Export trace for the latest receipt
trace_path = store.export_trace(session_id, receipt_path)
return session_id, messages, retrieved_view, ledger, receipt_path, receipt_path, trace_path, trace_path
def run_investor_demo(session_id: str, retrieval_k: int):
if not session_id:
session_id = new_session_id()
last = (session_id, [], "", "", "", None, "", None, "", None)
for step in INVESTOR_SCRIPT:
last = chat_turn(session_id, step, retrieval_k)
session_id = last[0]
return last
def verify_receipt_upload(file_obj) -> str:
if file_obj is None:
return "Upload a receipt JSON file."
with open(file_obj.name, "r", encoding="utf-8") as f:
data = json.load(f)
ok, msg = store.verify_receipt(data)
return f"{'✅' if ok else '❌'} {msg}"
def guardrail_tool_call(receipt_file, requested_action: str) -> str:
if receipt_file is None:
return "❌ Blocked: no receipt provided."
with open(receipt_file.name, "r", encoding="utf-8") as f:
receipt = json.load(f)
ok, msg = store.verify_receipt(receipt)
if not ok:
return f"❌ Blocked: receipt failed verification. Reason: {msg}"
# Allowed actions are deliberately boring in the demo.
# The point is the gate, not the tool.
return f"✅ Allowed: receipt verified. Executed action: {requested_action}"
def export_audit_pack(session_id: str):
if not session_id:
session_id = new_session_id()
path = store.export_audit_pack(session_id)
return path, path
def api_playground(session_id: str, action: str, payload_json: str, retrieval_k: int):
if not session_id:
session_id = new_session_id()
payload = {}
if payload_json and payload_json.strip():
payload = json.loads(payload_json)
if action == "memory.write":
role = payload.get("role", "user")
text = payload.get("text", "")
ev = store.append_event(session_id, role, text)
return session_id, json.dumps({"ok": True, "event": ev}, indent=2, ensure_ascii=False)
if action == "memory.search":
q = payload.get("query", "")
k = int(payload.get("k", retrieval_k))
hits = store.search_lexical(session_id, q, k=k)
out = {
"ok": True,
"hits": [{
"event_id": h.event_id, "seq": h.seq, "role": h.role, "text": h.text,
"score": h.score, "digest": h.digest, "chain_hash": h.chain_hash
} for h in hits]
}
return session_id, json.dumps(out, indent=2, ensure_ascii=False)
if action == "receipt.verify":
receipt = payload.get("receipt", {})
ok, msg = store.verify_receipt(receipt)
return session_id, json.dumps({"ok": ok, "message": msg}, indent=2, ensure_ascii=False)
if action == "audit.export":
path = store.export_audit_pack(session_id)
return session_id, json.dumps({"ok": True, "audit_pack_path": path}, indent=2, ensure_ascii=False)
return session_id, json.dumps({"ok": False, "error": "Unknown action"}, indent=2, ensure_ascii=False)
def token_savings(n: int, B: int, M: int, W: int) -> List[List[Any]]:
n = int(n)
B = int(B)
M = int(M)
W = int(W)
baseline = n * B + M * n * (n + 1) // 2
aifs = n * (B + W)
red = 0.0 if baseline == 0 else (1.0 - (aifs / baseline))
return [["turns", n],
["baseline_total_tokens", baseline],
["budgeted_total_tokens", aifs],
["reduction_percent", round(100.0 * red, 2)]]
def fill_example(selected: str) -> str:
return selected or ""
PITCH_MD = """
# RFTSystems TrustStack Console
I don’t do “trust me”. I do receipts.
This Space demonstrates an agent memory layer that is **durable, searchable, and provable**:
- Append-only session ledger (JSONL)
- SQLite FTS retrieval (fast lexical recall)
- Hash-chain integrity (tamper-evident)
- Per-turn “Memory Receipt” (what was retrieved + hashes)
- Receipt verification (pass/fail)
- Receipt-gated tool execution (guardrails)
- Trace export (what influenced what)
- Audit pack export (ZIP)
If you build agents for real users, receipts are what turns “memory” into infrastructure.
"""
HOW_TO_MD = """
## Quick demo
1) Click **Run Investor Demo**
2) Download the last receipt JSON
3) Upload it into **Verify Receipt**
4) Try the **Guardrails** tab: tool execution only passes with a verified receipt
5) Export an **Audit Pack** ZIP
## What this proves
- Memory persistence is not the hard part
- Retrieval is not the hard part
- **Integrity + evidence** is the hard part
"""
with gr.Blocks(title="RFTSystems TrustStack Console") as demo:
gr.Markdown(PITCH_MD)
with gr.Row():
session_id = gr.Textbox(label="Session ID", value=new_session_id())
retrieval_k = gr.Slider(1, 20, value=8, step=1, label="Retrieval K")
with gr.Tabs():
with gr.Tab("Investor Demo"):
gr.Markdown("One click. Full story: storage → retrieval → override → receipt → verification → trace → audit.")
run_demo_btn = gr.Button("Run Investor Demo", variant="primary")
demo_chat = gr.Chatbot(label="Demo Conversation", height=320)
demo_retrieved = gr.Textbox(label="Retrieved memory slices", lines=8)
demo_ledger = gr.Textbox(label="Ledger (hash-chained)", lines=12)
demo_receipt_path = gr.Textbox(label="Last receipt path (server)", lines=1)
demo_receipt_file = gr.File(label="Download last receipt JSON")
demo_trace_path = gr.Textbox(label="Last trace path (server)", lines=1)
demo_trace_file = gr.File(label="Download last trace JSON")
run_demo_btn.click(
run_investor_demo,
inputs=[session_id, retrieval_k],
outputs=[session_id, demo_chat, demo_retrieved, demo_ledger,
demo_receipt_path, demo_receipt_file, demo_trace_path, demo_trace_file],
)
with gr.Tab("Chat"):
chatbot = gr.Chatbot(label="Conversation", height=320)
with gr.Row():
example_pick = gr.Dropdown(label="Example prompts", choices=EXAMPLE_PROMPTS, value=EXAMPLE_PROMPTS[0])
use_example = gr.Button("Use Example", variant="secondary")
user_msg = gr.Textbox(label="Message")
send = gr.Button("Send", variant="primary")
retrieved_out = gr.Textbox(label="Retrieved memory slices", lines=8)
ledger_out = gr.Textbox(label="Ledger (hash-chained)", lines=12)
receipt_path = gr.Textbox(label="Last receipt path (server)", lines=1)
receipt_file = gr.File(label="Download last receipt JSON")
trace_path = gr.Textbox(label="Last trace path (server)", lines=1)
trace_file = gr.File(label="Download last trace JSON")
use_example.click(fill_example, inputs=[example_pick], outputs=[user_msg])
send.click(
chat_turn,
inputs=[session_id, user_msg, retrieval_k],
outputs=[session_id, chatbot, retrieved_out, ledger_out, receipt_path, receipt_file, trace_path, trace_file],
)
with gr.Tab("Verify Receipt"):
receipt_upload = gr.File(label="Upload receipt JSON")
verify_btn = gr.Button("Verify", variant="primary")
verify_out = gr.Textbox(label="Verification result")
verify_btn.click(verify_receipt_upload, inputs=[receipt_upload], outputs=[verify_out])
with gr.Tab("Guardrails"):
gr.Markdown("Tool execution is blocked unless the receipt verifies.")
receipt_for_tool = gr.File(label="Provide a receipt JSON")
action = gr.Dropdown(label="Requested action", choices=[
"tool.send_email_simulated",
"tool.export_customer_data_simulated",
"tool.trigger_purchase_simulated",
"tool.write_file_simulated",
], value="tool.write_file_simulated")
run_tool_btn = gr.Button("Attempt tool call", variant="primary")
tool_out = gr.Textbox(label="Result", lines=4)
run_tool_btn.click(guardrail_tool_call, inputs=[receipt_for_tool, action], outputs=[tool_out])
with gr.Tab("Audit Pack"):
gr.Markdown("One-click export: ledger + receipts + integrity summary.")
export_btn = gr.Button("Export Audit Pack ZIP", variant="primary")
audit_path = gr.Textbox(label="Audit pack path (server)", lines=1)
audit_file = gr.File(label="Download audit pack ZIP")
export_btn.click(export_audit_pack, inputs=[session_id], outputs=[audit_path, audit_file])
with gr.Tab("API Playground"):
gr.Markdown("API-style calls for the demo. JSON in, JSON out.")
action2 = gr.Dropdown(label="Action", choices=["memory.write", "memory.search", "receipt.verify", "audit.export"], value="memory.search")
payload = gr.Textbox(label="JSON payload", lines=10, value=json.dumps({"query": "Nova", "k": 8}, indent=2))
call_btn = gr.Button("Call", variant="primary")
api_out = gr.Textbox(label="Response", lines=14)
call_btn.click(api_playground, inputs=[session_id, action2, payload, retrieval_k], outputs=[session_id, api_out])
with gr.Tab("Token Budget"):
gr.Markdown("Why fixed retrieval budgets win as sessions grow.")
n = gr.Slider(1, 500, value=50, step=1, label="Turns (n)")
B = gr.Slider(0, 2000, value=500, step=50, label="Base framing tokens per call (B)")
M = gr.Slider(0, 2000, value=650, step=50, label="Avg tokens added per turn to history (M)")
W = gr.Slider(0, 8000, value=2000, step=100, label="Fixed retrieval budget tokens per call (W)")
calc = gr.Button("Compute", variant="primary")
tbl = gr.Dataframe(headers=["metric", "value"], datatype=["str", "str"], row_count=4, col_count=2)
calc.click(lambda nn, bb, mm, ww: token_savings(nn, bb, mm, ww), inputs=[n, B, M, W], outputs=[tbl])
with gr.Tab("How to Use"):
gr.Markdown(HOW_TO_MD)
demo.launch()