Building an AI search and knowledge retrieval workflow from scratch: a nanny-level tutorial

🛒 A complete practical guide for research, consulting, product and academic groups, from registration tools to knowledge accumulation.

Tutorial objectives and applicable readers

This tutorial will help you build a daily operation system for "AI search and knowledge retrieval" from scratch: from account registration, retrieval questioning skills, academic literature retrieval, to storing results into a personal knowledge base. No technical background is required, and it is applicable to research, consulting, product, operation, and academic scenarios. The entire process can be completed in about 60 minutes.

1. Preparation Checklist

  • [ ] Sign up for Perplexity (conversational search, free to get started)
  • [ ] Register Kimi (kimi.moonshot.cn, long text and Chinese search)
  • [ ] Register for Secret Tower AI search (structured search and library)
  • [ ] (optional) Register for Consensus or Elicit (academic literature search)
  • [ ] Prepare a note-taking tool (Notion/Feishu Documents/Obsidian can be used for knowledge accumulation)
  • [ ] Identify the first topic to be investigated (it is recommended to choose a real problem directly related to your work)

2. Account and environment preparation

  1. Open the Perplexity official website to complete registration and login.
  2. Open Kimi and log in, and become familiar with the "Upload File/Long Article Reading" entrance location.
  3. Open the Secret Tower AI search and experience switching between the three search modes of "webpage/library/academic".
  4. Create a "Search Item" page template in the note-taking tool, with suggested fields: topic, list of sub-questions, summary of conclusions, list of sources, items to be verified, and review.

3. Five-step method for retrieval and questioning

Good search questions determine the quality of the answers. Use the following five-step method to turn fuzzy requirements into searchable questions:

  1. Determine the subject: Clearly identify who and what the research object is (company/technology/policy/population).
  2. Determine the question: What do you want it to answer (current situation/comparison/impact/prediction/feasibility).
  3. Defined scope: Limit industries, regions, and time windows (such as "2024-2025 Chinese market").
  4. Determine the level of evidence: ordinary web pages, industry reports, or academic papers.
  5. Specify the output format: Require tables/key points/with sources.

Example comparison:

  • Bad: How about checking AI search for me
  • Good: Compare the functions and charging differences of the three AI search tools Perplexity, Kimi, and Secret Tower AI Search in 2025, output a comparison table, and attach a source link to each conclusion

4. Use Perplexity for conversational search

  1. Enter the sub-questions into Perplexity one by one and add unified requirements:
Please list your information sources item by item (with links), and mark the release time of each item; content whose source cannot be confirmed is marked separately with [to be verified].
  1. When comparison is required, explicitly require table output:
Please use the table to compare the differences between A and B in terms of price, functions, and applicable groups, and attach links to each source.
  1. Questioning and digging: Continue to ask questions about interesting clues in the answers, for example:
Which report is the specific data source you mentioned? Who is the publishing organization? Please give the original link.
  1. Copy each result to the "Conclusion Summary" and "Source List" fields of the note.

5. Use Kimi/Secret Tower to process long texts and Chinese materials

  1. When encountering a long report, PDF or long web page, upload the file to Kimi or paste the link and enter:
Please read the full text and output: 1) Core conclusions (within 5 items) 2) Key data table 3) Parts related to my topic 4) Unclarified but noteworthy doubts in the original text

  1. Use MITA AI Search to do Chinese structured search: switch to "Library" mode to search for reports/document materials, and switch to "Academic" mode to search for papers.

  2. When there is too much information, let AI do two rounds of screening:

First round: Screen out 10 items directly related to the topic from the above materials; Second round: Write a summary of each of these 10 items and mark the credibility (high/medium/low).

6. Use Consensus/Elicit for academic evidence retrieval

When the question involves "whether it is effective, what is the impact, and whether the evidence is sufficient", return to the academic literature:

  1. Open Consensus and search with English keywords (such as effect of remote work on productivity).
  2. Check the distribution of literature conclusions (support/contra/neutral), and click on highly relevant papers to read abstracts.
  3. Use Elicit to perform a review-style search: enter a research question and let it extract the conclusions, methods, and sample sizes of multiple papers and generate a comparison table.
  4. Note: Academic conclusions have scope of application. Be sure to read the limitations and applicable groups marked in the abstract to avoid over-interpretation.

7. Knowledge accumulation

The completion of retrieval does not mean the end. The results are precipitated into reusable assets:

  1. Fill in the note template: topic, summary of conclusions, source list (including links and dates), and items to be verified.
  2. Mark the credibility of the conclusion: high (consistent from multiple sources + official), medium (single source or second-hand), low (AI speculation without source).
  3. Regularly (weekly) merge similar topics and refine them into "theme cards" for reuse by the team.
  4. If the company has a lot of internal data, it can be evaluated to connect to the Glean enterprise search tool to turn the internal documents into a conversational search portal (IT cooperation is required).

8. Verification method

  • Source traceability: spot-check 5 conclusions and find the original source within 2 minutes each
  • Accuracy: Key data is consistent with the official page
  • Timeliness: the release time of the search results is within the time window you require
  • Precipitation rate: After the completion of the survey, the note template completion rate is ≥90%
  • Efficiency: The second search time for the same topic is reduced by more than 50% compared with the first time

9. Frequently Asked Questions and Troubleshooting (FAQ)

  1. There is no source in the AI search answer? It is mandatory to "attach a source link to each conclusion" in the question. Content without a source is considered to be verified.
  2. The result is too old? Add time limits when asking questions, such as "Only search information since January 2024"; the key data is subject to the official real-time page.
  3. Chinese and English materials are mixed together and it’s very confusing? Check one round for each language: use Kimi/Secret Tower for Chinese, use Perplexity/Academic Tools for English, and then merge them.
  4. Are the citations retrieved compiled by AI? Academic tools may provide documents that do not exist. Be sure to click on the original text to check the volume number/DOI. If you cannot find it, delete it.
  5. The free quota is not enough? Use the free file for daily searches, and use the paid file for periods when important topics are concentrated; store high-frequency questions into knowledge cards to reduce repeated searches.
  6. The company intranet information cannot be retrieved? Public tools cannot retrieve intranet content, so you need to access the company's internal search tools and complete permission configuration.

10. Advanced expansion

  • API Automation: Integrate the retrieval link into the automated process to achieve "question → automatic retrieval → abstract → entry into the knowledge base"
  • Team Knowledge Base: Unify the precipitation caliber and build a team search entrance with permissions
  • Search template library: Create a set of standard question templates for each high-frequency research scenario (competitive product analysis, industry scanning, technology selection)
  • Linked Writing: The search results are directly used as material input for the AI writing pipeline to improve content production efficiency.

11. Practical example: complete research on a topic

Take the topic "Comparison of Mainstream AI Search Tools in 2025" as an example to string together the entire process:

  1. Split the question: Divide it into 4 sub-questions - ① What are the mainstream tools? ② Their respective functions and charges; ③ What scenarios are they suitable for? ④ User reputation and shortcomings.
  2. Perplexity Preliminary Screening: Enter 4 sub-questions respectively, add "output comparison table with source link", and get a draft answer with source.
  3. Chinese supplement: Use Secret Tower AI to search for industry reports in the "library" mode, and use Kimi to read two long reports to extract key points.
  4. Academic supplement: For questions such as "Does AI search improve retrieval efficiency?", use Consensus to check the paper evidence and record samples and conclusions.
  5. Cross-validation: Compare the pricing pages of each official website with third-party reviews, and mark the consistency and inconsistencies.
  6. Precipitation: Fill in the conclusion, source list, items to be verified according to the note template, mark the credibility, and generate the "AI Search Tool Comparison Card".

After completion, a random check of all 5 conclusions can be traced back to the source within 2 minutes, which is considered qualified. Save the sub-question template and question phrase as a "retrieval template" and reuse it directly next time for similar topics.

User Reviews

  • Loading reviews...