Maintained by deepset

Integration: Parallel

Use the Parallel Search API and Parallel's web-research model in Haystack pipelines.

Authors

deepset

GitHub Repo PyPI Package

Overview
Installation
Usage
License

Overview

Parallel builds web research infrastructure for AI agents. The parallel-haystack package brings it to Haystack through two components:

ParallelWebSearch — web search through the Parallel Search API, returning LLM-optimized excerpts as Haystack Document objects.
ParallelChatGenerator — chat completion through the Parallel Responses API, where every answer is grounded in live web research with citations.

For more information about the Parallel APIs, see the Parallel docs.

To follow along with this guide, you’ll need a Parallel API key from the Parallel platform. Add it as an environment variable, PARALLEL_API_KEY.

Installation

pip install parallel-haystack

Usage

You can use the Parallel components as standalone components or in Haystack pipelines.

Use Parallel Web Search (Search API)

ParallelWebSearch returns three outputs: documents (excerpts with title, url and, where available, publish_date metadata), links (the result URLs), and session_id.

import os

from haystack.utils import Secret
from haystack_integrations.components.websearch.parallel import ParallelWebSearch

os.environ["PARALLEL_API_KEY"] = "YOUR_PARALLEL_API_KEY"

websearch = ParallelWebSearch(
    api_key=Secret.from_env_var("PARALLEL_API_KEY"),
    top_k=5,
)
result = websearch.run(query="What is Haystack by deepset?")

print(result["documents"])
print(result["links"])

The Search API offers four modes — turbo, fast, basic and advanced — in increasing order of latency and quality. Pass the mode and any other Search API parameter through search_params, either at initialization or per run() call:

from haystack.utils import Secret
from haystack_integrations.components.websearch.parallel import ParallelWebSearch

websearch = ParallelWebSearch(
    api_key=Secret.from_env_var("PARALLEL_API_KEY"),
    top_k=5,
    search_params={
        "mode": "turbo",
        "advanced_settings": {"source_policy": {"include_domains": ["arxiv.org"]}},
    },
)
result = websearch.run(query="Latest retrieval-augmented generation research")

for document in result["documents"]:
    print(document.meta["title"], document.meta["url"])

Searches that belong to the same task can share a session. Pass the session_id returned by one search back into the next one to get better contextual results:

from haystack.utils import Secret
from haystack_integrations.components.websearch.parallel import ParallelWebSearch

websearch = ParallelWebSearch(api_key=Secret.from_env_var("PARALLEL_API_KEY"))

first = websearch.run(query="What is Haystack by deepset?")
second = websearch.run(
    query="Who maintains Haystack?",
    search_params={"session_id": first["session_id"]},
)

print(second["links"])

Use Parallel Web Research (Responses API)

ParallelChatGenerator runs live web research for every call and returns an answer with citations. The reasoning.effort parameter selects the research tier: low (~5-10s), medium (~15-20s, the default) or high (~30-60s).

import os

from haystack.dataclasses import ChatMessage
from haystack_integrations.components.generators.parallel import ParallelChatGenerator

os.environ["PARALLEL_API_KEY"] = "YOUR_PARALLEL_API_KEY"

client = ParallelChatGenerator(generation_kwargs={"reasoning": {"effort": "low"}})
response = client.run(
    messages=[ChatMessage.from_user("What did Parallel Web Systems announce this year?")]
)

print(response["replies"][0].text)

Because web grounding is built into the model, tool calling and sampling parameters such as tools, temperature and top_p are accepted for SDK compatibility but have no effect on the response; the component logs a warning when it sees them.

Use Parallel in a pipeline

This RAG pipeline searches the web with ParallelWebSearch and answers from the retrieved excerpts.

import os

from haystack import Pipeline
from haystack.components.builders import ChatPromptBuilder
from haystack.dataclasses import ChatMessage
from haystack.utils import Secret
from haystack_integrations.components.generators.parallel import ParallelChatGenerator
from haystack_integrations.components.websearch.parallel import ParallelWebSearch

os.environ["PARALLEL_API_KEY"] = "YOUR_PARALLEL_API_KEY"

prompt_builder = ChatPromptBuilder(
    template=[
        ChatMessage.from_system("You are a helpful assistant."),
        ChatMessage.from_user(
            "Given the information below:\n"
            "{% for document in documents %}{{ document.content }}\n{% endfor %}\n"
            "Answer the following question: {{ query }}.\nAnswer:"
        ),
    ],
    required_variables=["query", "documents"],
)

pipe = Pipeline()
pipe.add_component("search", ParallelWebSearch(api_key=Secret.from_env_var("PARALLEL_API_KEY"), top_k=3))
pipe.add_component("prompt_builder", prompt_builder)
pipe.add_component("llm", ParallelChatGenerator(generation_kwargs={"reasoning": {"effort": "low"}}))

pipe.connect("search.documents", "prompt_builder.documents")
pipe.connect("prompt_builder.prompt", "llm.messages")

query = "What is Haystack by deepset?"
result = pipe.run(data={"search": {"query": query}, "prompt_builder": {"query": query}})

print(result["llm"]["replies"][0].text)

License

parallel-haystack is distributed under the terms of the Apache-2.0 license.