标签：#Python

GitHub热门项目 | | Stars: 346 | 71 stars today | 语言: Python

GitHub热门项目 | 🕵️‍♂️ (2-in-1) Email & Username OSINT suite for deep data extraction just from a single Email/Username. Analyzes 455+ actively maintained scan vectors (175+ email / 280+ username) for security research, investigations, and digital footprinting. | Stars: 3,481 | 39 stars today | 语言: Python

A few weekends ago I got annoyed at typing prompts into a terminal and decided the fix was, obviously, to control my AI agent with hand gestures instead. This is the story of building that, and the two hours I lost fighting a GPU crash that had nothing to do with my code. The idea: a webcam watches your hand, MediaPipe tracks the landmarks, and three gestures map to three actions on an Anthropic-powered coding agent. Pinch (thumb and index touching) - the agent writes code Spinning your index finger in a circle - the agent brainstorms an idea Two fingers "running" up and down - it runs whatever code it just wrote No keyboard. No prompt box. Just your hand in front of a webcam, like you're a conductor telling an orchestra what to play. The MediaPipe detour I started with MediaPipe's newer Tasks API (HandLandmarker), because it's the one all the docs point you to now. It crashed immediately on my Mac with a Metal/GPU service error, even when I forced it onto the CPU delegate. Spent way too long assuming it was my setup before realizing the new API just doesn't play nice with this machine. Switched to the legacy mp.solutions.hands API, pinned to mediapipe==0.10.21, and the problem vanished. Sometimes the fix for a shiny new API is to not use it yet. Gestures are messier than they sound Detecting "pinch" is easy: measure the distance between thumb and index tip, threshold it, done. The other two took more work. "Running" fingers needed the vertical oscillation of the index and middle fingertips, counted by sign crossings, so it doesn't false trigger on a hand that's just drifting. "Spinning" tracks the index fingertip's trajectory and accumulates the signed angle around a center point, so a real circle reads differently than a shaky hand. Both run on a rolling 1.5 second buffer of landmarks, edge triggered so a gesture fires once, not once per frame. Letting the agent run its own code, unsandboxed, on purpose The runner executes whatever the agent wrote as a subprocess

I created this article for the purpose of entering the All Things Agentic Hackathon. TL;DR — An ADK output_key writes into the session of the agent that declares it. In-process that session is shared, so it looks like state flows. Across a RemoteA2aAgent hop it is the worker's session, and it never comes back. Nothing raises. Nothing warns. Every local run and every CI job exercises the working topology, so the failure is invisible to an offline test suite by construction — including at 100% coverage. The system that passed Bastion is a three-agent access-governance fleet built with Google ADK and A2A. An Orchestrator owns investigation state, an Access Auditor reads production IAM through a read-only identity, and a model-free Escalation Agent delivers validated count-only reviews. The local graph passed its configured core statement and branch coverage gate. Every branch, every seam. Then the same graph was split across deployed A2A workers, and an assumption that looked natural in-process became false. The boundary we had not modeled In-process, the previous step's result is simply there : # The Auditor declares output_key; the Orchestrator reads it back. report = ctx . session . state . get ( AUDIT_FINDINGS_KEY ) Deploy the same sequence and only the construction changes. The graph is identical: RemoteA2aAgent ( name = " access_auditor " , agent_card = card_url ( auditor , " access_auditor " ), description = " Reads the live IAM policy and flags anomalies. Read-only. " , httpx_client = private_a2a_client ( auditor ), a2a_request_meta_provider = _forward_investigation , ) output_key still writes. It writes into the worker's session, which never crosses back. The deployed Orchestrator saw an empty state key while every local run and every test saw a populated one. Observed 2026-08-22: the Auditor completed a full sub-trail, and the next step then refused with "returned no structured report." No exception at the boundary. No warning at construction. The run still r

The AI Wrote the Diff. The Tests Wrote the Verdict. AI refactor suggestions are hypotheses. Not facts. A free coding model rewrites your messy legacy function. The diff looks clean. CI stays green. Then a customer hits an edge case you forgot. This article shows a small workflow. Characterize legacy behavior first. Let the model propose a refactor. Run the same tests against both versions. The verdict: safe or not safe. Disclosure: This article was prepared as part of MonkeyCode's product outreach. Why Characterization Comes First Legacy code has no spec. The only reliable spec is current behavior. Even bugs are behavior. If your refactor changes a bug, you need to know. A characterization test records inputs and outputs. It does not judge right or wrong. It freezes the current contract. After freezing, every difference becomes visible. Step 1: Capture Real Inputs and Outputs Pick one messy function. I used a shipping calculator. Nested conditionals, magic numbers, zero tests. Write a probe script. Call the function with realistic cases. Save outputs as JSON. import json from legacy import calculate_shipping cases = [ { ' items ' : [{ ' weight ' : 2.0 , ' qty ' : 3 }], ' region ' : ' US ' }, { ' items ' : [{ ' weight ' : 0.5 , ' qty ' : 10 }], ' region ' : ' EU ' }, { ' items ' : [{ ' weight ' : 0.2 , ' qty ' : 1 }], ' region ' : ' US ' }, { ' items ' : [{ ' weight ' : 5.0 , ' qty ' : 2 }], ' region ' : ' JP ' }, ] for c in cases : result = calculate_shipping ( c [ ' items ' ], c [ ' region ' ]) print ( json . dumps ({ ' input ' : c , ' output ' : result })) Save output to captured.json . That becomes ground truth. Step 2: Ask the Model for a Refactor MonkeyCode's free model access lets me prompt from the CLI. I gave the model one strict instruction: keep behavior identical. Refactor calculate_shipping into smaller functions. Do NOT change edge cases. Do NOT change rounding. Extract private helpers only. The model returned a diff. It split the function into three he

By Michael Nocito , data analyst · Published August 8, 2026 By the end of this page you can load a CSV into pandas, find out in twenty seconds what type every column became, stop the identifier columns losing their leading zeros, get dates read the way they were written, and turn a money column that arrived as text into numbers. It is about twenty-five minutes, and every output below was produced by running the code. Here is what to do today, the moment after you first load a file. Run df.dtypes . Not df.head() , which shows you what the values look like, but dtypes , which shows you what they are. A column of identifiers that says int64 has already lost its leading zeros, and a money column that says object or str is text that will refuse to add up. The short version: read_csv reads characters and guesses a type per column. The guess is usually right, it is silent when it is wrong, and four arguments replace guessing with instruction. The same characters becoming two different values is the idea, so it gets the picture. The original carries a diagram here. In words: On the left, a strip of five small square boxes holds one character each, reading zero, eight, zero, five, three, as the characters appear in the file. Two arrows branch out from that strip. The upper arrow leads to a strip of five boxes in which the first box is empty, crossed through and outlined in amber, while the remaining four hold eight, zero, five and three; the leading character has been discarded. The lower arrow leads to a strip of five boxes holding zero, eight, zero, five and three, identical to the original, outlined in blue. Both destinations came from the same source strip, and only one of them still contains everything the file did. Every output on this page is real. Run on pandas 3.0.2 against a small CSV built to contain the four problems every real export has: an identifier with leading zeros, ambiguous dates, a text marker for missing values, and money with a thousands separator. If

About a year ago, I was building a terminal-based workflow manager called Glyph.Flow. It was mostly a learning project. I wanted to understand Python better, experiment with Textual, think about commands, state, configuration, logging, and all the small architectural decisions that suddenly appear when a script stops being a script. Somewhere between then and now, the workflows became a little more real. For my Master's thesis, I built a data pipeline to construct and process a cross-national research database from multiple sources. It had a clear purpose: take heterogeneous input data, transform it consistently, validate important assumptions, and produce the dataset I needed for the analysis. And it worked. But this is no longer enough. I am not rebuilding it because the original system failed. I am rebuilding it because the question changed: My Master's thesis needed a pipeline. My PhD will need research infrastructure. And I am slowly discovering that these are not the same thing. A pipeline can be finished There is something comfortable about building software for a well-defined research project. You know the research question. You know most of the variables you need. You know which datasets are involved. You can define the transformations, produce the outputs, validate them, run the analysis, and eventually say: Done. Of course, research is never really that clean. Data sources change. Weird edge cases appear. A country disappears from one dataset. Another source changes a variable name. An indicator turns out to mean something slightly different than you thought. But there is still a boundary around the problem. A PhD changes that boundary. Now I have to think about a system that may need to survive several years of research, new questions I have not formulated yet, datasets I have not discovered yet, and methodological decisions I will probably reconsider more than once. Suddenly, "Does it work?" becomes a surprisingly weak design criterion. The more useful

By Michael Nocito , data analyst · Published August 8, 2026 By the end of this page you can turn transactions into a monthly series, add period-on-period change and a cumulative total, get a share-of-total column, smooth a noisy line, and run all of it separately for every group. It is about twenty-five minutes, and every number below came out of running the code. Here is what to do today, on the series you already have. Count its rows against the number of periods in your date range. If your data covers January to May and the series has four rows, a period produced nothing, it never became a row, and every change figure after the gap is comparing the wrong pair. The short version: pct_change() divides each value by the one in the row above; cumsum() adds everything up to and including the current row. Both trust the rows you gave them to be the periods you meant. What happens when the previous period is zero is the idea, so it gets the picture. The original carries a diagram here. In words: Three bar positions stand on a baseline, labelled Mar, Apr and May. The March position holds a tall bar and the May position holds a slightly shorter tall bar. The April position holds no bar at all; there is only a short flat mark sitting on the baseline where a bar would start, drawn in amber to show a value of zero. An arc runs from the top of the March bar down to the April mark, and the figure minus one hundred percent is printed on it, which is a perfectly ordinary answer. A second arc runs from the April mark up to the top of the May bar, and the symbol printed on that one is not a percentage at all but the sideways figure eight that means infinity. The picture shows that a fall to nothing has an answer and a rise from nothing does not. Every number on this page is real. The sixteen-row orders table used across this whole set of guides, run in pandas 3.0.2. It runs from 5 January to 25 May 2026 and contains no April orders at all, which is not staged for this page; it is

Python PostgreSQL with asyncpg: Async Database Operations asyncpg is the fastest PostgreSQL driver for Python — pure asyncio, no thread overhead, and up to 3× faster than psycopg2 on typical workloads. It is the go-to choice for any async Python backend. Installation pip install asyncpg # PostgreSQL server must already be running Connect and Create a Pool import asyncio import asyncpg from datetime import datetime DATABASE_URL = " postgresql://user:password@localhost:5432/mydb " async def create_pool () -> asyncpg . Pool : pool = await asyncpg . create_pool ( DATABASE_URL , min_size = 2 , max_size = 10 , command_timeout = 30 , server_settings = { " application_name " : " myapp " }, ) print ( " Pool created. " ) return pool Schema Setup CREATE_TABLES = """ CREATE TABLE IF NOT EXISTS users ( id BIGSERIAL PRIMARY KEY, username TEXT NOT NULL UNIQUE, email TEXT NOT NULL UNIQUE, created_at TIMESTAMPTZ NOT NULL DEFAULT now() ); CREATE TABLE IF NOT EXISTS posts ( id BIGSERIAL PRIMARY KEY, user_id BIGINT NOT NULL REFERENCES users(id) ON DELETE CASCADE, title TEXT NOT NULL, body TEXT NOT NULL DEFAULT '' , published BOOLEAN NOT NULL DEFAULT FALSE, created_at TIMESTAMPTZ NOT NULL DEFAULT now() ); CREATE INDEX IF NOT EXISTS idx_posts_user ON posts(user_id); CREATE INDEX IF NOT EXISTS idx_posts_created ON posts(created_at DESC); """ async def setup_schema ( pool : asyncpg . Pool ) -> None : async with pool . acquire () as conn : await conn . execute ( CREATE_TABLES ) print ( " Schema ready. " ) INSERT — Adding Records async def create_user ( pool : asyncpg . Pool , username : str , email : str ) -> int : async with pool . acquire () as conn : row = await conn . fetchrow ( """ INSERT INTO users (username, email) VALUES ($1, $2) ON CONFLICT (username) DO UPDATE SET email = EXCLUDED.email RETURNING id, created_at """ , username , email , ) return row [ " id " ] async def create_post ( pool : asyncpg . Pool , user_id : int , title : str , body : str , published : bool = False , ) ->

By Michael Nocito , data analyst · Published August 8, 2026 By the end of this page you can attach columns from one DataFrame to another on a shared key, choose the right how for the question, see at a glance which rows failed to match, and catch the failure that quietly inflates every total in the frame. It is about twenty-five minutes, and every output below was produced by running the code. Here is what to do today, on every merge you write. Print the row count immediately before and immediately after it. A left merge must not change the row count, and if it did, the right-hand table has the key more than once and your totals have just gone up. The short version: merge pairs rows from two frames wherever their keys match, and the number of rows that come out depends on how many times each key appears on each side. One key twice on the right is the idea, so it gets the picture. The original carries a diagram here. In words: On the left a single row is drawn as a wide box, holding the key Desk and the value 880. To its right stands a small lookup table with two rows, and both of those rows carry the same key, Desk. Two lines run from the single left-hand row, one to each of the two matching lookup rows, so the one row is paired twice. On the far right the result is drawn as two separate output rows, and both of them contain Desk and 880; the value 880 is ringed in amber in each of them to show that it is the same original figure appearing twice. One row went in and two came out, without anything being added to the left-hand table. Every output on this page is real. Sixteen orders totalling 9,890 and a three-row product table, the same tables used across this whole set of guides, merged in pandas 3.0.2 with the results copied back. If you know SQL joins , this is the same operation with different words, and the two failure modes are identical. 1. merge in one line Two frames, one shared column, one call. orders.merge(products, on="product", how="left") order_id prod

Google Flights has a date grid with a fare for every departure day, and a "track prices" toggle that emails you when its pick of dates moves. Both are fine for one trip. Neither gives you the table: every day, the fare, the airline and stops behind it, in rows you can sort, keep and put a threshold on. If your dates are flexible and you want the cheapest day to fly as data — or airfare price tracking that runs every morning — you need rows. This is how to get one row per departure day from Google Flights as JSON, with no API key (there is no public Google Flights API), and how to turn it into a flight price alert. 1. One request, one row per day The Flight Price Tracker on Apify takes routes, a first departure day and a window length. For each day it searches Google Flights, keeps that day's cheapest itinerary and ranks the days. A 30-day window on one route is at most 30 fare rows plus a free status row. curl -X POST "https://api.apify.com/v2/acts/kestrel~flight-price-tracker/run-sync-get-dataset-items?token= $APIFY_TOKEN " \ -H "Content-Type: application/json" \ -d '{"routes": ["LIS-LHR"], "departDate": "2026-10-05", "days": 30, "adults": 1, "currency": "USD", "market": "us"}' A fare row: { "type" : "fare" , "route" : "LIS-LHR" , "trip" : "one_way" , "depart_date" : "2026-10-05" , "return_date" : null , "seat" : "economy" , "adults" : 1 , "currency" : "USD" , "price" : 127 , "price_display" : "127 US dollars" , "airline" : "Tap Air Portugal" , "stops" : 0 , "depart_time" : "8:00 PM" , "arrive_time" : "10:55 PM" , "duration" : "2 hr 55 min" , "duration_minutes" : 175 , "layovers" : null , "co2_kg" : 123 , "itineraries_seen" : 12 , "cheapest_in_window" : true , "rank_in_window" : 1 , "google_url" : "https://www.google.com/travel/flights?tfs=..." , "fetched_at" : "2026-08-29T06:25:14+00:00" } cheapest_in_window is true on exactly one day per route; rank_in_window orders the rest. The free status row repeats the headline as cheapest and cheapest_date , with days_searc

Catching Prompt Injection Before It Enters a Trusted Knowledge Base AI agents increasingly consume knowledge from sources they did not author and cannot independently verify: a PDF policy document, a scraped web page, a spreadsheet exported from another team's system. The prevailing approach — extract the text, write it into a knowledge base or context window, let the agent treat it as fact — has an underexamined weakness. Extraction tools capture everything present in a source document, including content a human reviewer would never see. The Mechanism Several ordinary, well-documented features of common file formats allow text to be present in a document while remaining invisible to anyone reading it normally: A PDF can render text in a rendering mode that instructs viewers not to display it, or set its fill color identical to the page background. A Word document has an explicit "hidden" attribute on any run of text, independent of color or size. A PowerPoint file's speaker notes are parsed by most extraction tools but never appear to an audience watching the presentation. A spreadsheet can mark entire rows, columns, or sheets as hidden, or attach a comment to a cell that is invisible unless hovered. An HTML page can hide an element from a browser's rendering entirely via a handful of standard CSS properties. None of these are obscure edge cases. They are common, legitimate formatting features, used constantly for entirely benign reasons — a hidden helper column in a spreadsheet, a private note to a presenter, draft text a Word user hid rather than deleted. The problem is not that these features exist; it is that an extraction pipeline has no reason to distinguish "this text is legitimate content" from "this text was deliberately hidden" unless something is specifically checking for the difference. Why This Matters for AI Pipelines Specifically If an attacker can place text anywhere in this chain — inside a PDF a company will later ingest, inside a web page a scrap

Pulled live from leetcode.com/problemset/?difficulty=Hard on 29 Aug 2026 (895 hard problems in the Algorithms list). "First 30" = the 30 lowest problem numbers. Everything below is Python 3 . How to use this Open the problem on LeetCode and make sure the language selector says Python3 . Select all the text in the code editor and delete it. Paste the block below in its place — each block already contains the class Solution signature LeetCode generated for that problem, plus any commented-out ListNode / TreeNode header. Press Submit . Do not add import statements or redefine ListNode / TreeNode — LeetCode injects typing.List , typing.Optional , heapq , math.gcd and the node classes automatically. The blocks are written to rely on exactly that. Verification Every solution was executed locally against an independent brute-force reference on randomised and edge-case inputs ( 4,637 assertions, all passing ), then stress-tested at each problem's documented maximum input size ( 31/31 within budget ). Two real defects were found and fixed during that pass — see the notes on #127 and #149. 4. Median of Two Sorted Arrays https://leetcode.com/problems/median-of-two-sorted-arrays/ Approach. Binary search on the cut position of the shorter array. O(log(min(m,n))) , O(1) space. Constraints (from the problem page). nums1.length == m nums2.length == n 0 <= m <= 1000 0 <= n <= 1000 1 <= m + n <= 2000 -10 6 <= nums1[i], nums2[i] <= 10 6 class Solution : def findMedianSortedArrays ( self , nums1 : List [ int ], nums2 : List [ int ]) -> float : # Binary search on the shorter array's cut position. O(log(min(m, n))). if len ( nums1 ) > len ( nums2 ): nums1 , nums2 = nums2 , nums1 m , n = len ( nums1 ), len ( nums2 ) lo , hi = 0 , m total = ( m + n + 1 ) // 2 while lo <= hi : i = ( lo + hi ) // 2 # take i elements from nums1 j = total - i # take j elements from nums2 l1 = nums1 [ i - 1 ] if i > 0 else float ( ' -inf ' ) r1 = nums1 [ i ] if i < m else float ( ' inf ' ) l2 = nums2 [ j - 1 ]

I created this content for the purposes of entering the All Things Agentic Hackathon. The problem that started it A doctor gets about eight minutes with a patient and, for anyone with a real history, forty pages of scattered records — lab reports, discharge notes, and pharmacy bills from three different clinics. So the history is effectively invisible at the exact moment it matters most. And when the visit ends, nothing follows up: the six-month course lapses at week five, the recheck never gets booked. I wanted to build an agent that closes that loop — one that reads the mess, decides urgency in a way a clinician can actually trust, and handles the follow-up on its own. That became CareLoop , my entry for the All Things Agentic Hackathon (Taskmaster track), built on Gemini, the Google Agent Development Kit (ADK), Cloud Run, and Firestore. The one principle I wouldn't compromise on Rules decide, AI explains. The temptation with an LLM is to let it do everything — including deciding whether a chest-pain patient is urgent. I refused to do that. In CareLoop, a deterministic engine owns every clinical decision: a weighted symptom score plus a red-flag override sets the triage level and routing. It is fully auditable, and it returns byte-identical output on the same input every single time. The LLM's job is strictly language: Reading unstructured documents into a fixed schema — I call it "Gemini extracts, rules merge." Writing the structured result into a plain-language brief a clinician can skim in ten seconds. No language model is ever in the decision path. When a judge asks "why was this Critical?", the answer is a score breakdown they can inspect — not a model's say-so. That single decision shaped the whole architecture. What it actually does CareLoop runs the full loop end to end: Ingest & compact — it reads a patient's documents and merges them into one structured ledger: allergies, chronic conditions, active medications, and lab trends over time. Instead of pushin

Most model benchmarks tell you how smart the model is on the first attempt, which is almost never the problem in production. The real problem is what happens on the 120th attempt, when the same kind of input shows up again and nobody is watching. I spent 48 hours running the same classification task against a free model on a free server, and the drift taught me more than accuracy ever did. The Setup I'd Run Again The workload was dull on purpose: ten support tickets, three labels, one prompt template. Every hour the job asked the model to classify one ticket and logged the raw output, so each ticket appeared about twelve times. It was not a benchmark of intelligence; it was a probe of stability, and stability is what automation actually needs. I ran the whole thing on MonkeyCode's free server option, using the free model access for inference, because a cheap long-running job is exactly the scenario that setup is for. Disclosure: This article was prepared as part of MonkeyCode's product outreach. The rest is about what the probe caught, not about quotas or latency, so treat my numbers as one operator's field notes. The Probe Code (Steal This) A probe is only honest if it writes down everything, including the outputs you didn't ask for. The script below hashes every response, tries to parse a label, and appends one JSON line per run, so nothing interesting ever gets lost. import hashlib , json , time LOG_PATH = " drift.jsonl " LABELS = ( " bug " , " feature " , " question " ) def stable_hash ( text ): return hashlib . sha256 ( text . strip (). encode ()). hexdigest ()[: 12 ] def parse_label ( raw ): # Accepts JSON or plain prose; returns None when the format is unknown. try : return json . loads ( raw ). get ( " label " ) except json . JSONDecodeError : found = [ label for label in LABELS if label in raw ] return found [ 0 ] if found else None def record_run ( run_id , ticket_id , raw , expected ): entry = { " run " : run_id , " ticket " : ticket_id , " hash " : stabl

Google Hotels already compares every booking site for a hotel and a stay — Booking.com, Expedia, Agoda, Hotels.com and the hotel's own site. It has a "track prices" button too, but it emails you on its own terms, picks the sources, and keeps the history. If you want the numbers — for a trip, a rate parity check, or a price history chart — you need them as rows. This is how to get Google Hotels prices for exact dates as JSON, without a Google API key (there is no public Google Hotels API for reading prices; the official Hotel APIs are feeds for hotels sending prices to Google), and how to turn that into daily hotel price tracking. 1. One request, every booking site's rate The Google Hotels Prices Scraper on Apify takes a place search or a list of hotels, a stay, occupancy and currency, and returns three row types: hotel (lowest nightly rate + stay total), offer (each source's rate, free‑cancellation flag, deep link) and status . You pay per priced row; sold‑out hotels and empty searches are free. curl -X POST "https://api.apify.com/v2/acts/kestrel~google-hotels-prices/run-sync-get-dataset-items?token= $APIFY_TOKEN " \ -H "Content-Type: application/json" \ -d '{"queries": ["hotels in Lisbon"], "checkIn": "2026-10-03", "checkOut": "2026-10-06", "adults": 2, "currency": "USD", "maxHotels": 20}' A hotel row looks like this: { "type" : "hotel" , "name" : "The Central House Lisbon Baixa" , "check_in" : "2026-10-03" , "check_out" : "2026-10-06" , "nights" : 3 , "nightly" : 81.81 , "nightly_display" : "$82" , "total" : 245 , "stars" : 2 , "rating" : 4.3 , "reviews" : 727 , "deal" : "19% less than usual" , "entity_id" : "ChkIg-b2ismUj7M1Gg0vZy8xMWg3MThreGg1EAE" , "google_url" : "https://www.google.com/travel/hotels/entity/ChkI…" } and an offer row (with "includeOffers": true ): { "type" : "offer" , "name" : "Hyatt Regency Lisbon" , "source" : "Booking.com" , "official" : false , "nightly" : 569.35 , "total" : 1708.05 , "free_cancel" : true , "free_cancel_until" : "Oct 1" , "p

Building Next-Gen Agentic Architectures: From Local RAG to Sandboxed Execution and BigQuery MCP The 3 AM production fire revealed a harsh truth: modern agentic systems often collapse under their own weight. A single agent processing 10K RAG queries OOM-killed an 8GB cloud instance. The culprit was not the workload but the infrastructure: @pinecone-client/vecdb with 47 transitive dependencies bloat memory with unquantized float32 embeddings. The solution was 200 lines of Python using sqlite3 , array , and heapq , with bounded queues and race condition resilience. This is the story of how we replaced dependency bloat with surgical precision. The Dependency Problem Agentic systems today face three critical bottlenecks: Vector Search : Libraries like faiss-cpu (12MB) combined with pg-vector (synchronous disk I/O) block the event loop, creating latency spikes. BigQuery : The @google-cloud/bigquery client (12MB) plus grpcio (5MB) leaks file descriptors, hitting Linux's default 1024 soft limit. Sandboxing : Docker containers consume 500MB+ per instance, making them impractical for memory-constrained environments. The root cause is always the same: unbounded resource consumption. 1M vectors at 768 dimensions in float32 consumes 3GB of memory. Synchronous I/O stalls the event loop. Unmanaged connections leak file descriptors. The Zero-Bloat RAG Engine The solution begins with a fundamental shift: replace heavy dependencies with lightweight, audited code. Our LocalRAG implementation demonstrates this approach: import sqlite3 import array import heapq import json import threading from typing import List , Tuple , Optional class LocalRAG : def __init__ ( self , db_path : str , dim : int = 768 , max_vectors : int = 1_000_000 ): self . dim = dim self . max_vectors = max_vectors self . lock = threading . Lock () self . conn = sqlite3 . connect ( db_path , isolation_level = None , check_same_thread = False ) # Enable WAL mode for concurrent reads/writes self . conn . execute ( " PR

Nobody likes phone trees. "Press 1 for billing, press 2 for support." Miss an option? Start over. It is friction at its worst. The voice-ivr-with-agent-backend example replaces that with a natural language conversation. Callers just say what they need, and the app routes them to the right department. Code: https://github.com/team-telnyx/telnyx-code-examples/tree/main/voice-ivr-with-agent-backend What it builds A Python/Flask app that handles inbound calls with a conversational IVR: Inbound Call -> answer with Call Control -> look up menu config from KV -> LLM generates a dynamic greeting -> gather(speech) — caller says what they need -> LLM routes intent to a department -> transfer call The core primitives The app combines four Telnyx primitives: Call Control : answer() , speak() , gather_using_speech() , transfer() AI Inference : telnyx.ai.openai.chat.completions.create() for greetings and intent routing KV store : menu config per phone number (business name, departments, transfer numbers, keywords) Agent state machine : an IVRAgent class that tracks call state, turn count, and retry logic Dynamic greeting via LLM Instead of a hardcoded "Press 1 for billing," the app generates a conversational greeting from the KV config: def generate_dynamic_menu_prompt ( menu_config : dict ) -> str : departments = menu_config . get ( " departments " , []) dept_list = " \n " . join ( f " - { d [ ' name ' ] } : { d [ ' description ' ] } " for d in departments ) return ( f " You are an IVR assistant for { menu_config [ ' business_name ' ] } . " f " Available departments: \n { dept_list } \n\n " f " Greet the caller briefly and ask how you can help. " f " Keep it conversational and under 2 sentences. " ) The LLM generates the greeting through the OpenAI-compatible Telnyx Inference binding. If it fails, the app falls back to a static greeting from the KV config. Intent routing via LLM When the caller speaks, the transcription is passed to route_intent_with_llm . The LLM is instructed

TL;DR: Running high-frequency autonomous AI agent loops on commercial LLM APIs at scale is economically unsustainable and introduces unpredictable latency spikes. This production guide details how we deployed a self-hosted inference cluster using vLLM (v0.6+) , EAGLE-3 speculative decoding , PagedAttention v2 , and Automatic Prefix Caching (APC) on cloud GPUs (RunPod/Vast.ai), achieving a sub-180ms Time-To-First-Token (TTFT) , 118 tokens/sec throughput , and cutting inference costs by 45–74% . 1. The Economic & Latency Bottleneck of Agentic Loops When building 24/7 autonomous daemon agents , LangGraph multi-agent state machines , or LLM-driven NPC game loops , the computational profile differs fundamentally from human chatbot interactions: Massive Request Volume: A single complex agent decision cycle frequently executes 5 to 25 LLM calls across intent classification, tool schema validation, reflection loops, and output formatting. Repeated Prefix Redundancy: 80–90% of prompt tokens consist of identical system instructions, persona framing, and MCP (Model Context Protocol) tool definitions. Strict Latency Budgets: Real-time simulations and game loops cannot tolerate 800ms–1500ms commercial API network roundtrips. Commercial Closed APIs (GPT-4o / Claude 3.5 Sonnet) ├── Prefill: Paid per-token on every single cyclic call ├── Network Roundtrip: 250ms - 600ms latency overhead └── Cost at 50,000 daily agent iterations: $1,200 - $3,500 / month Self-Hosted vLLM Cluster (RTX 4090 / A100 on RunPod) ├── Automatic Prefix Caching (APC): Reuses KV-cache (120ms -> 12ms prefill) ├── Speculative Decoding (EAGLE-3): 2.1x generation throughput └── Fixed Infrastructure Cost: $245 - $480 / month (Flat, unlimited tokens) 2. Deep Dive: The vLLM Memory & Scheduling Architecture PagedAttention: Eliminating KV-Cache Fragmentation Standard PyTorch/HuggingFace transformer implementations allocate static KV-cache tensors sized for max_sequence_length . Because 95% of queries generate far fewer

GitHub热门项目 | Python toolkit for quantitative finance | Stars: 12,772 | 25 stars today | 语言: Python