<spanid="arch-overview-prompt-handling"></span><h1>Prompts<a@click.prevent="window.navigator.clipboard.writeText($el.href); $el.setAttribute('data-tooltip', 'Copied!'); setTimeout(() => $el.setAttribute('data-tooltip', 'Copy link to this element'), 2000)"aria-label="Copy link to this element"class="headerlink"data-tooltip="Copy link to this element"href="#prompts"><svgheight="1em"viewbox="0 0 24 24"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M3.9 12c0-1.71 1.39-3.1 3.1-3.1h4V7H7c-2.76 0-5 2.24-5 5s2.24 5 5 5h4v-1.9H7c-1.71 0-3.1-1.39-3.1-3.1zM8 13h8v-2H8v2zm9-6h-4v1.9h4c1.71 0 3.1 1.39 3.1 3.1s-1.39 3.1-3.1 3.1h-4V17h4c2.76 0 5-2.24 5-5s-2.24-5-5-5z"></path></svg></a></h1>
<p>Arch’s <strong>prompt handler</strong> subsystem interacts with the <strong>model subsytem</strong> through Envoy’s cluster manager system to ensure robust, resilient and fault-tolerant experience in managing incoming prompts.</p>
<divclass="admonition seealso">
<pclass="admonition-title">See also</p>
<p>Read more about the <aclass="reference internal"href="model_serving.html#model-serving"><spanclass="std std-ref">model subsystem</span></a> and how the LLMs are hosted in Arch.</p>
<h2>Messages<a@click.prevent="window.navigator.clipboard.writeText($el.href); $el.setAttribute('data-tooltip', 'Copied!'); setTimeout(() => $el.setAttribute('data-tooltip', 'Copy link to this element'), 2000)"aria-label="Copy link to this element"class="headerlink"data-tooltip="Copy link to this element"href="#messages"x-intersect.margin.0%.0%.-70%.0%="activeSection ='#messages'"><svgheight="1em"viewbox="0 0 24 24"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M3.9 12c0-1.71 1.39-3.1 3.1-3.1h4V7H7c-2.76 0-5 2.24-5 5s2.24 5 5 5h4v-1.9H7c-1.71 0-3.1-1.39-3.1-3.1zM8 13h8v-2H8v2zm9-6h-4v1.9h4c1.71 0 3.1 1.39 3.1 3.1s-1.39 3.1-3.1 3.1h-4V17h4c2.76 0 5-2.24 5-5s-2.24-5-5-5z"></path></svg></a></h2>
<p>Arch accepts messages directly from the body of the HTTP request in a format that follows the <aclass="reference external"href="https://huggingface.co/docs/text-generation-inference/en/messages_api"rel="nofollow noopener">Hugging Face Messages API<svgfill="currentColor"height="1em"stroke="none"viewbox="0 96 960 960"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M188 868q-11-11-11-28t11-28l436-436H400q-17 0-28.5-11.5T360 336q0-17 11.5-28.5T400 296h320q17 0 28.5 11.5T760 336v320q0 17-11.5 28.5T720 696q-17 0-28.5-11.5T680 656V432L244 868q-11 11-28 11t-28-11Z"></path></svg></a>.
This design allows developers to pass a list of messages, where each message is represented as a dictionary
containing two key-value pairs:</p>
<blockquote>
<div><ulclass="simple">
<li><p><strong>Role</strong>: Defines the role of the message sender, such as “user” or “assistant”.</p></li>
<li><p><strong>Content</strong>: Contains the actual text of the message.</p></li>
<p>Arch is engineered with <aclass="reference internal"href="../../guides/prompt_guard.html#prompt-guard"><spanclass="std std-ref">Arch-Guard</span></a>, an industry leading safety layer, powered by a
compact and high-performimg LLM that monitors incoming prompts to detect and reject jailbreak attempts -
ensuring that unauthorized or harmful behaviors are intercepted early in the process.</p>
<p>To add jailbreak guardrails, see example below:</p>
<divclass="code-block-caption"><spanclass="caption-text">Example Configuration</span><a@click.prevent="window.navigator.clipboard.writeText($el.href); $el.setAttribute('data-tooltip', 'Copied!'); setTimeout(() => $el.setAttribute('data-tooltip', 'Copy link to this element'), 2000)"aria-label="Copy link to this element"class="headerlink"data-tooltip="Copy link to this element"href="#id1"><svgheight="1em"viewbox="0 0 24 24"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M3.9 12c0-1.71 1.39-3.1 3.1-3.1h4V7H7c-2.76 0-5 2.24-5 5s2.24 5 5 5h4v-1.9H7c-1.71 0-3.1-1.39-3.1-3.1zM8 13h8v-2H8v2zm9-6h-4v1.9h4c1.71 0 3.1 1.39 3.1 3.1s-1.39 3.1-3.1 3.1h-4V17h4c2.76 0 5-2.24 5-5s-2.24-5-5-5z"></path></svg></a></div>
</span><spanid="line-6"><spanclass="linenos"> 6</span><spanclass="w"></span><spanclass="c1"># Defines how Arch should parse the content from application/json or text/pain Content-type in the http request</span>
</span><spanid="line-9"><spanclass="linenos"> 9</span><spanclass="c1"># Centralized way to manage LLMs, manage keys, retry logic, failover and limits in a central way</span>
</span><spanid="line-19"><spanclass="linenos">19</span><spanclass="nt">system_prompt</span><spanclass="p">:</span><spanclass="w"></span><spanclass="l l-Scalar l-Scalar-Plain">You are a network assistant that just offers facts; not advice on manufacturers or purchasing decisions.</span>
</mark></span><spanid="line-25"><mark><spanclass="linenos">25</span><spanclass="w"></span><spanclass="nt">message</span><spanclass="p">:</span><spanclass="w"></span><spanclass="l l-Scalar l-Scalar-Plain">Looks like you're curious about my abilities, but I can only provide assistance within my programmed parameters.</span>
<h2>Prompt Targets<a@click.prevent="window.navigator.clipboard.writeText($el.href); $el.setAttribute('data-tooltip', 'Copied!'); setTimeout(() => $el.setAttribute('data-tooltip', 'Copy link to this element'), 2000)"aria-label="Copy link to this element"class="headerlink"data-tooltip="Copy link to this element"href="#prompt-targets"x-intersect.margin.0%.0%.-70%.0%="activeSection ='#prompt-targets'"><svgheight="1em"viewbox="0 0 24 24"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M3.9 12c0-1.71 1.39-3.1 3.1-3.1h4V7H7c-2.76 0-5 2.24-5 5s2.24 5 5 5h4v-1.9H7c-1.71 0-3.1-1.39-3.1-3.1zM8 13h8v-2H8v2zm9-6h-4v1.9h4c1.71 0 3.1 1.39 3.1 3.1s-1.39 3.1-3.1 3.1h-4V17h4c2.76 0 5-2.24 5-5s-2.24-5-5-5z"></path></svg></a></h2>
<p>Once a prompt passes any configured guardrail checks, Arch processes the contents of the incoming conversation
and identifies where to forwad the conversation to via its <codeclass="docutils literal notranslate"><spanclass="pre">prompt_targets</span></code> primitve. Prompt targets are endpoints
that receive prompts that are processed by Arch. For example, Arch enriches incoming prompts with metadata like knowing
when a user’s intent has changed so that you can build faster, more accurate RAG apps.</p>
<p>Configuring <codeclass="docutils literal notranslate"><spanclass="pre">prompt_targets</span></code> is simple. See example below:</p>
<divclass="code-block-caption"><spanclass="caption-text">Example Configuration</span><a@click.prevent="window.navigator.clipboard.writeText($el.href); $el.setAttribute('data-tooltip', 'Copied!'); setTimeout(() => $el.setAttribute('data-tooltip', 'Copy link to this element'), 2000)"aria-label="Copy link to this element"class="headerlink"data-tooltip="Copy link to this element"href="#id2"><svgheight="1em"viewbox="0 0 24 24"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M3.9 12c0-1.71 1.39-3.1 3.1-3.1h4V7H7c-2.76 0-5 2.24-5 5s2.24 5 5 5h4v-1.9H7c-1.71 0-3.1-1.39-3.1-3.1zM8 13h8v-2H8v2zm9-6h-4v1.9h4c1.71 0 3.1 1.39 3.1 3.1s-1.39 3.1-3.1 3.1h-4V17h4c2.76 0 5-2.24 5-5s-2.24-5-5-5z"></path></svg></a></div>
</span><spanid="line-6"><spanclass="linenos"> 6</span><spanclass="w"></span><spanclass="c1"># Defines how Arch should parse the content from application/json or text/pain Content-type in the http request</span>
</span><spanid="line-9"><spanclass="linenos"> 9</span><spanclass="c1"># Centralized way to manage LLMs, manage keys, retry logic, failover and limits in a central way</span>
</span><spanid="line-19"><spanclass="linenos">19</span><spanclass="nt">system_prompt</span><spanclass="p">:</span><spanclass="w"></span><spanclass="l l-Scalar l-Scalar-Plain">You are a network assistant that just offers facts; not advice on manufacturers or purchasing decisions.</span>
</span><spanid="line-25"><spanclass="linenos">25</span><spanclass="w"></span><spanclass="nt">message</span><spanclass="p">:</span><spanclass="w"></span><spanclass="l l-Scalar l-Scalar-Plain">Looks like you're curious about my abilities, but I can only provide assistance within my programmed parameters.</span>
</span><spanid="line-30"><spanclass="linenos">30</span><spanclass="w"></span><spanclass="nt">description</span><spanclass="p">:</span><spanclass="w"></span><spanclass="l l-Scalar l-Scalar-Plain">handel all scenarios that are question and answer in nature. Like summarization, information extraction, etc.</span>
</span><spanid="line-34"><spanclass="linenos">34</span><spanclass="w"></span><spanclass="c1"># Arch uses the default LLM and treats the response from the endpoint as the prompt to send to the LLM</span>
</span><spanid="line-36"><spanclass="linenos">36</span><spanclass="w"></span><spanclass="c1"># override system prompt for this prompt target</span>
</span><spanid="line-37"><spanclass="linenos">37</span><spanclass="w"></span><spanclass="nt">system_prompt</span><spanclass="p">:</span><spanclass="w"></span><spanclass="l l-Scalar l-Scalar-Plain">You are a helpful information extraction assistant. Use the information that is provided to you.</span>
</mark></span><spanid="line-40"><mark><spanclass="linenos">40</span><spanclass="w"></span><spanclass="nt">description</span><spanclass="p">:</span><spanclass="w"></span><spanclass="l l-Scalar l-Scalar-Plain">Reboot a specific network device</span>
</mark></span><spanid="line-47"><mark><spanclass="linenos">47</span><spanclass="w"></span><spanclass="nt">description</span><spanclass="p">:</span><spanclass="w"></span><spanclass="l l-Scalar l-Scalar-Plain">Identifier of the network device to reboot.</span>
</mark></span><spanid="line-51"><mark><spanclass="linenos">51</span><spanclass="w"></span><spanclass="nt">description</span><spanclass="p">:</span><spanclass="w"></span><spanclass="l l-Scalar l-Scalar-Plain">Confirmation flag to proceed with reboot.</span>
</span><spanid="line-60"><spanclass="linenos">60</span><spanclass="c1"># Arch creates a round-robin load balancing between different endpoints, managed via the cluster subsystem.</span>
</span><spanid="line-63"><spanclass="linenos">63</span><spanclass="w"></span><spanclass="c1"># value could be ip address or a hostname with port</span>
</span><spanid="line-64"><spanclass="linenos">64</span><spanclass="w"></span><spanclass="c1"># this could also be a list of endpoints for load balancing</span>
</span><spanid="line-65"><spanclass="linenos">65</span><spanclass="w"></span><spanclass="c1"># for example endpoint: [ ip1:port, ip2:port ]</span>
</span><spanid="line-67"><spanclass="linenos">67</span><spanclass="w"></span><spanclass="c1"># max time to wait for a connection to be established</span>
<p>Check <aclass="reference internal"href="../prompt_target.html#prompt-target"><spanclass="std std-ref">Prompt Target</span></a> for more details!</p>
<h3>Intent Detection and Prompt Matching:<a@click.prevent="window.navigator.clipboard.writeText($el.href); $el.setAttribute('data-tooltip', 'Copied!'); setTimeout(() => $el.setAttribute('data-tooltip', 'Copy link to this element'), 2000)"aria-label="Copy link to this element"class="headerlink"data-tooltip="Copy link to this element"href="#intent-detection-and-prompt-matching"x-intersect.margin.0%.0%.-70%.0%="activeSection ='#intent-detection-and-prompt-matching'"><svgheight="1em"viewbox="0 0 24 24"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M3.9 12c0-1.71 1.39-3.1 3.1-3.1h4V7H7c-2.76 0-5 2.24-5 5s2.24 5 5 5h4v-1.9H7c-1.71 0-3.1-1.39-3.1-3.1zM8 13h8v-2H8v2zm9-6h-4v1.9h4c1.71 0 3.1 1.39 3.1 3.1s-1.39 3.1-3.1 3.1h-4V17h4c2.76 0 5-2.24 5-5s-2.24-5-5-5z"></path></svg></a></h3>
<p>Arch uses fast Natural Language Inference (NLI) and embedding approaches to first detect the intent of each
incoming prompt. This intent detection phase analyzes the prompt’s content and matches it against predefined
prompt targets, ensuring that each prompt is forwarded to the most appropriate endpoint. Arch’s intent
detection framework considers both the name and description of each prompt target, and uses a composite matching
score between an NLI and cosine similarity to enchance accuracy in forwarding decisions.</p>
<ulclass="simple">
<li><p><strong>Embeddings</strong>: By embedding the prompt and comparing it to known target vectors, Arch effectively identifies
the closest match, ensuring that the prompt is handled by the correct downstream service.</p></li>
<li><p><strong>NLI</strong>: NLI techniques further refine the matching process by evaluating the semantic alignment between the
prompt and potential targets.</p></li>
</ul>
</section>
<sectionid="agentic-apps-via-prompt-targets">
<h3>Agentic Apps via Prompt Targets<a@click.prevent="window.navigator.clipboard.writeText($el.href); $el.setAttribute('data-tooltip', 'Copied!'); setTimeout(() => $el.setAttribute('data-tooltip', 'Copy link to this element'), 2000)"aria-label="Copy link to this element"class="headerlink"data-tooltip="Copy link to this element"href="#agentic-apps-via-prompt-targets"x-intersect.margin.0%.0%.-70%.0%="activeSection ='#agentic-apps-via-prompt-targets'"><svgheight="1em"viewbox="0 0 24 24"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M3.9 12c0-1.71 1.39-3.1 3.1-3.1h4V7H7c-2.76 0-5 2.24-5 5s2.24 5 5 5h4v-1.9H7c-1.71 0-3.1-1.39-3.1-3.1zM8 13h8v-2H8v2zm9-6h-4v1.9h4c1.71 0 3.1 1.39 3.1 3.1s-1.39 3.1-3.1 3.1h-4V17h4c2.76 0 5-2.24 5-5s-2.24-5-5-5z"></path></svg></a></h3>
<p>To support agentic apps, like scheduling travel plans or sharing comments on a document - via prompts, Arch uses
its function calling abilities to extract critical information from the incoming prompt (or a set of prompts)
needed by a downstream backend API or function call before calling it directly. For more details on how you can
build agentic applications using Arch, see our full guide <aclass="reference internal"href="../../build_with_arch/agent.html#arch-agent-guide"><spanclass="std std-ref">here</span></a>:</p>
<divclass="admonition note">
<pclass="admonition-title">Note</p>
<p>Arch <aclass="reference internal"href="../../guides/function_calling.html#function-calling"><spanclass="std std-ref">Arch-Function</span></a> is the dedicated agentic model engineered in Arch to extract information from
a (set of) prompts and executes necessary backend API calls. This allows for efficient handling of agentic tasks,
such as scheduling data retrieval, by dynamically interacting with backend services. Arch-Function is a flagship 1.3
billion parameter model that matches performance with frontier models like Claude Sonnet 3.5 ang GPT-4, while
being 100x cheaper ($0.05M/token hosted) and 10x faster (p50 latencies of 200ms).</p>
</div>
</section>
</section>
<sectionid="prompting-llms">
<h2>Prompting LLMs<a@click.prevent="window.navigator.clipboard.writeText($el.href); $el.setAttribute('data-tooltip', 'Copied!'); setTimeout(() => $el.setAttribute('data-tooltip', 'Copy link to this element'), 2000)"aria-label="Copy link to this element"class="headerlink"data-tooltip="Copy link to this element"href="#prompting-llms"x-intersect.margin.0%.0%.-70%.0%="activeSection ='#prompting-llms'"><svgheight="1em"viewbox="0 0 24 24"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M3.9 12c0-1.71 1.39-3.1 3.1-3.1h4V7H7c-2.76 0-5 2.24-5 5s2.24 5 5 5h4v-1.9H7c-1.71 0-3.1-1.39-3.1-3.1zM8 13h8v-2H8v2zm9-6h-4v1.9h4c1.71 0 3.1 1.39 3.1 3.1s-1.39 3.1-3.1 3.1h-4V17h4c2.76 0 5-2.24 5-5s-2.24-5-5-5z"></path></svg></a></h2>
<p>Arch is a single piece of software that is designed to manage both ingress and egress prompt traffic, drawing its
distributed proxy nature from the robust <aclass="reference external"href="https://envoyproxy.io"rel="nofollow noopener">Envoy<svgfill="currentColor"height="1em"stroke="none"viewbox="0 96 960 960"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M188 868q-11-11-11-28t11-28l436-436H400q-17 0-28.5-11.5T360 336q0-17 11.5-28.5T400 296h320q17 0 28.5 11.5T760 336v320q0 17-11.5 28.5T720 696q-17 0-28.5-11.5T680 656V432L244 868q-11 11-28 11t-28-11Z"></path></svg></a>. This makes it extremely efficient and capable
of handling upstream connections to LLMs. If your application is originating code to an API-based LLM, simply use
the OpenAI client and configure it with Arch. By sending traffic through Arch, you can propagate traces, manage and monitor
traffic, apply rate limits, and utilize a large set of traffic management capabilities in a centralized way.</p>
<divclass="admonition attention">
<pclass="admonition-title">Attention</p>
<p>When you start Arch, it automatically creates a listener port for egress calls to upstream LLMs. This is based on the
<codeclass="docutils literal notranslate"><spanclass="pre">llm_providers</span></code> configuration section in the <codeclass="docutils literal notranslate"><spanclass="pre">prompt_config.yml</span></code> file. Arch binds itself to a local address such as
<h3>Example: Using OpenAI Client with Arch as an Egress Gateway<a@click.prevent="window.navigator.clipboard.writeText($el.href); $el.setAttribute('data-tooltip', 'Copied!'); setTimeout(() => $el.setAttribute('data-tooltip', 'Copy link to this element'), 2000)"aria-label="Copy link to this element"class="headerlink"data-tooltip="Copy link to this element"href="#example-using-openai-client-with-arch-as-an-egress-gateway"x-intersect.margin.0%.0%.-70%.0%="activeSection ='#example-using-openai-client-with-arch-as-an-egress-gateway'"><svgheight="1em"viewbox="0 0 24 24"width="1em"xmlns="http://www.w3.org/2000/svg"><pathd="M3.9 12c0-1.71 1.39-3.1 3.1-3.1h4V7H7c-2.76 0-5 2.24-5 5s2.24 5 5 5h4v-1.9H7c-1.71 0-3.1-1.39-3.1-3.1zM8 13h8v-2H8v2zm9-6h-4v1.9h4c1.71 0 3.1 1.39 3.1 3.1s-1.39 3.1-3.1 3.1h-4V17h4c2.76 0 5-2.24 5-5s-2.24-5-5-5z"></path></svg></a></h3>
<p>In these examples, the OpenAI client is used to send traffic directly through the Arch egress proxy to the LLM of your choice, such as OpenAI.
The OpenAI client is configured to route traffic via Arch by setting the proxy to <codeclass="docutils literal notranslate"><spanclass="pre">127.0.0.1:51001</span></code>, assuming Arch is running locally and bound to that address and port.
This setup allows you to take advantage of Arch’s advanced traffic management features while interacting with LLM APIs like OpenAI.</p>
<li><a:data-current="activeSection === '#example-using-openai-client-with-arch-as-an-egress-gateway'"class="reference internal"href="#example-using-openai-client-with-arch-as-an-egress-gateway">Example: Using OpenAI Client with Arch as an Egress Gateway</a></li>
