Configuration

The config.py enables you to customize GPT Researcher to your specific needs and preferences.

Thanks to our amazing community and contributions, GPT Researcher supports multiple LLMs and Retrievers. In addition, GPT Researcher can be tailored to various report formats (such as APA), word count, research iterations depth, etc.

GPT Researcher defaults to our recommended suite of integrations: OpenAI for LLM calls and Tavily API for retrieving real-time web information.

Choosing an LLM

GPT Researcher's writer turns scraped sources into a report, so how faithfully a model sticks to its sources matters more than general benchmark scores. The table below comes from Vectara's Hallucination Leaderboard, which has each model summarize 7,700+ articles (news, science, medicine, law, business and more) using only the source text, then checks every summary for claims the source doesn't support with the HHEM-2.3 evaluation model.

A selection of models you can configure in GPT Researcher, ordered by hallucination rate (lower is better). Data as of September 22, 2026; see the leaderboard for all 100+ models.

Model	Provider	Hallucination rate	Factual consistency	Answer rate
GPT-5.4 nano	OpenAI	3.1%	96.9%	100.0%
Gemini 2.5 Flash-Lite	Google	3.3%	96.7%	99.5%
Llama 3.3 70B	Meta	4.1%	95.9%	99.5%
Mistral Large (24.11)	Mistral	4.5%	95.5%	99.9%
GPT-5.4 mini	OpenAI	5.5%	94.5%	100.0%
GPT-4.1	OpenAI	5.6%	94.4%	99.9%
DeepSeek V3.2	DeepSeek	6.3%	93.7%	92.6%
GPT-6 Sol	OpenAI	6.5%	93.5%	100.0%
GPT-5.4	OpenAI	7.0%	93.0%	99.9%
Gemini 2.5 Pro	Google	7.0%	93.0%	99.1%
Gemini 2.5 Flash	Google	7.8%	92.2%	99.0%
Llama 4 Maverick	Meta	8.2%	91.8%	100.0%
DeepSeek V4 Pro	DeepSeek	8.6%	91.4%	97.2%
GPT-5.5	OpenAI	9.3%	90.7%	100.0%
Claude Haiku 4.5	Anthropic	9.8%	90.2%	99.5%
Gemini 3.1 Pro (preview)	Google	10.4%	89.6%	99.4%
Claude Sonnet 4.6	Anthropic	10.6%	89.4%	99.9%
Qwen 3.5 Plus	Alibaba	10.7%	89.3%	99.8%
Kimi K2.6	Moonshot	10.8%	89.2%	99.7%
Claude Opus 4.7	Anthropic	12.0%	88.0%	98.0%
gpt-oss-120b	OpenAI	14.2%	85.8%	99.9%
Grok 4.1 Fast	xAI	17.8%	82.2%	98.5%

Hallucination rate: share of summaries containing a claim not supported by the source.
Answer rate: share of articles the model agreed to summarize.

GPT Researcher defaults to gpt-5.4 for SMART_LLM and STRATEGIC_LLM, and gpt-5.4-mini for FAST_LLM. This benchmark measures faithfulness to sources, not reasoning or writing quality: smaller models often rank well here partly because their summaries add less. Use it to narrow your choice, then compare reports on your own queries (see Testing your LLM).

The default config.py file can be found in /gpt_researcher/config/. It supports various options for customizing GPT Researcher to your needs. You can also include your own external JSON file config.json by adding the path in the config_path param. The config JSON should follow the format/keys in the default config. Below is a sample config.json file to help get you started:

{
  "RETRIEVER": "tavily",
  "EMBEDDING": "openai:text-embedding-3-small",
  "SIMILARITY_THRESHOLD": 0.42,
  "CONTEXT_FILTER": "auto",
  "FAST_LLM": "openai:gpt-5.4-mini",
  "SMART_LLM": "openai:gpt-5.4",
  "STRATEGIC_LLM": "openai:gpt-5.4",
  "LANGUAGE": "english",
  "CURATE_SOURCES": false,
  "FAST_TOKEN_LIMIT": 3000,
  "SMART_TOKEN_LIMIT": 6000,
  "STRATEGIC_TOKEN_LIMIT": 4000,
  "BROWSE_CHUNK_MAX_LENGTH": 8192,
  "SUMMARY_TOKEN_LIMIT": 700,
  "TEMPERATURE": 0.4,
  "DOC_PATH": "./my-docs",
  "REPORT_SOURCE": "web"
}

For example, to start GPT-Researcher and specify a specific config you would do this:

python gpt_researcher/main.py --config_path my_config.json

Please follow the config.py file for additional future support.

Below is a list of current supported options:

RETRIEVER: Search engine or research retriever used for retrieving sources. Defaults to tavily. Options include tavily, duckduckgo, bing, brave, google, searchapi, serper, serpapi, searx, arxiv, openalex, semantic_scholar, pubmed_central, exa, crw, groundroute, bocha, xquik, custom, and mcp. You can also combine retrievers with commas, such as tavily,openalex,semantic_scholar. Check here for supported retrievers
EMBEDDING: Embedding model. Defaults to openai:text-embedding-3-small. Options: ollama, huggingface, azure_openai, custom.
SIMILARITY_THRESHOLD: Threshold value for similarity comparison when processing documents. Defaults to 0.42.
CONTEXT_FILTER: How scraped content is filtered before it reaches the writer: auto (default; Jev when TYPESAFE_API_KEY is set, otherwise keyword ranking), jev, keyword, embeddings or none. No mode requires an embeddings provider except embeddings. See Context Filter.
FAST_LLM: Model name for fast LLM operations such summaries. Defaults to openai:gpt-5.4-mini.
SMART_LLM: Model name for smart operations like generating research reports and reasoning. Defaults to openai:gpt-5.4.
STRATEGIC_LLM: Model name for strategic operations like generating research plans and strategies. Defaults to openai:gpt-5.4.
LANGUAGE: Language to be used for the final research report. Defaults to english.
CURATE_SOURCES: Whether to curate sources for research. This step adds an LLM run which may increase costs and total run time but improves quality of source selection. Defaults to False.
FAST_TOKEN_LIMIT: Maximum token limit for fast LLM responses. Defaults to 3000.
SMART_TOKEN_LIMIT: Maximum token limit for smart LLM responses. Defaults to 6000.
STRATEGIC_TOKEN_LIMIT: Maximum token limit for strategic LLM responses. Defaults to 4000.

Recommended values for modern long-output models

The default token limits are calibrated for GPT-4o-class models (16k max output). For models with larger output capacity, increase these limits to avoid truncated reports:

Model family	Max output	Recommended SMART_TOKEN_LIMIT
GPT-4o / GPT-4.1	16k	8000
Claude Haiku 4.5	64k	16000
Claude Sonnet 4.6	64k	16000
Claude Opus 4.7	128k	32000
GPT-5 family	128k	32000

Apply proportional values to FAST_TOKEN_LIMIT and STRATEGIC_TOKEN_LIMIT if you use distinct models for those roles. The hard upper bound is 200k (sanity guard against typos).

BROWSE_CHUNK_MAX_LENGTH: Maximum length of text chunks to browse in web sources. Defaults to 8192.
SUMMARY_TOKEN_LIMIT: Maximum token limit for generating summaries. Defaults to 700.
TEMPERATURE: Sampling temperature for LLM responses, typically between 0 and 1. A higher value results in more randomness and creativity, while a lower value results in more focused and deterministic responses. Defaults to 0.4.
USER_AGENT: Custom User-Agent string for web crawling and web requests.
MAX_SEARCH_RESULTS_PER_QUERY: Maximum number of search results to retrieve per query. Defaults to 5.
MEMORY_BACKEND: Backend used for memory operations, such as local storage of temporary data. Defaults to local.
TOTAL_WORDS: Total word count limit for document generation or processing tasks. Defaults to 1200.
REPORT_FORMAT: Preferred format for report generation. Defaults to APA. Consider formats like MLA, CMS, Harvard style, IEEE, etc.
MAX_ITERATIONS: Maximum number of iterations for processes like query expansion or search refinement. Defaults to 3.
AGENT_ROLE: Role of the agent. This configures the behavior of specialized research agents. Defaults to None. When set, it activates role-specific prompting and techniques tailored to particular research domains.
MAX_SUBTOPICS: Maximum number of subtopics to generate or consider. Defaults to 3.
SCRAPER: Web scraper to use for gathering information. Defaults to bs (BeautifulSoup). You can also use newspaper.
MAX_SCRAPER_WORKERS: Maximum number of concurrent scraper workers per research. Defaults to 15.
REPORT_SOURCE: Source for the research report data. Defaults to web for online research. Can be set to doc for local document-based research. This determines where GPT Researcher gathers its primary information from.
DOC_PATH: Path to read and research local documents. Defaults to ./my-docs.
PROMPT_FAMILY: The family of prompts and prompt formatting to use. Defaults to prompting optimized for GPT models. See the full list of options in enum.py.
LLM_KWARGS: Json formatted dict of additional keyword args to be passed to the LLM provider class when instantiating it. This is primarily useful for clients like Ollama that allow for additional keyword arguments such as num_ctx that influence the inference calls.
EMBEDDING_KWARGS: Json formatted dict of additional keyword args to be passed to the embedding provider class when instantiating it.
DEEP_RESEARCH_BREADTH: Controls the breadth of deep research, defining how many parallel paths to explore. Defaults to 3.
DEEP_RESEARCH_DEPTH: Controls the depth of deep research, defining how many sequential searches to perform. Defaults to 2.
DEEP_RESEARCH_CONCURRENCY: Controls the concurrency level for deep research operations. Defaults to 4.
REASONING_EFFORT: Controls the reasoning effort of strategic models. Default to medium.

Deep Research Configuration

The deep research parameters allow you to fine-tune how GPT Researcher explores complex topics that require extensive knowledge gathering. These parameters work together to determine the thoroughness and efficiency of the research process:

DEEP_RESEARCH_BREADTH: Controls how many parallel research paths are explored simultaneously. A higher value (e.g., 5) causes the researcher to investigate more diverse subtopics at each step, resulting in broader coverage but potentially less focus on core themes. The default value of 3 provides a balanced approach between breadth and depth.
DEEP_RESEARCH_DEPTH: Determines how many sequential search iterations GPT Researcher performs for each research path. A higher value (e.g., 3-4) allows for following citation trails and diving deeper into specialized information, but increases research time substantially. The default value of 2 ensures reasonable depth while maintaining practical completion times.
DEEP_RESEARCH_CONCURRENCY: Sets how many concurrent operations can run during deep research. Higher values speed up the research process on capable systems but may increase API rate limit issues or resource consumption. The default value of 4 is suitable for most environments, but can be increased on systems with more resources or decreased if you experience performance issues.

For academic or highly specialized research, consider increasing both breadth and depth (e.g., BREADTH=4, DEPTH=3). For quick exploratory research, lower values (e.g., BREADTH=2, DEPTH=1) will provide faster results with less detail.

To change the default configurations, you can simply add env variables to your .env file as named above or export manually in your local project directory.

For example, to manually change the search engine and report format:

export RETRIEVER=bing
export REPORT_FORMAT=IEEE

For academic literature reviews, you can combine web and scholarly retrievers:

export RETRIEVER=tavily,openalex,semantic_scholar

Please note that you might need to export additional env vars and obtain API keys for other supported search retrievers and LLM providers. Please follow your console logs for further assistance. To learn more about additional LLM support you can check out the docs here.

Choosing an LLM​

Recommended values for modern long-output models​

Deep Research Configuration​

Choosing an LLM

Recommended values for modern long-output models

Deep Research Configuration