Development approach for a state-owned enterprise bidding and procurement document knowledge base agent: making bidding and procurement materials searchable, traceable, and auditable

People doing procurement and bidding work at state-owned enterprises have probably all had a similar experience: an audit team comes in and asks for the bid evaluation report of a project from three years ago. The person in charge has been replaced three times, the USB drive has been swapped five times, and finally, in a corner of the hard drive of a scrapped computer, a folder named "Final Version Really No More Changes" is dug out. Upon opening it, the reply record for the clarification letter is still missing. The project itself is perfectly clean, yet in the face of the four words "incomplete materials," one is left unable to explain oneself no matter how much one protests.

The difficulty of bidding document management lies not in "storage," but in whether they can be retrieved quickly, completely, and verifiably when needed.

This is not management laxity on the part of any single organization; rather, the category of bidding documents is inherently difficult to manage: there are many laws and regulations, project cycles are long, participants are diverse, and formats vary widely. More critically, it bears not merely the function of "storing" but serves as the underlying evidence chain for audit tracing, compliance proof, and experience reuse. DA/T 103-2024, the "Specification for Archiving Electronic Documents of Bidding and Tendering," has raised the four requirements of "authenticity, completeness, security, and usability" to an unprecedented height, effectively handing all state-owned enterprise procurement and bidding departments an exam paper they must catch.

The bidding document knowledge base agent solves precisely the questions on this exam paper. But the word "agent" easily leads people astray—they assume it is yet another chat window, answering one question at a time, bustling for three days before no one uses it. A bidding knowledge base that can truly run in a state-owned enterprise environment is developed precisely not starting from "conversation," but from "organizing the materials to the point where they can be trusted."

Step One: First Break "Materials" Down into "Retrievable Evidence"

The first step of a bidding knowledge base is not model selection, but governing regulation versions, project ledgers, and historical cases into something retrievable and citable.

What is special about bidding documents is that a single bidding document simultaneously overlays regulatory basis, project facts, time nodes, and approval traces. If one simply throws PDFs into a vector database for semantic retrieval, asking "What is the bid evaluation method for this project?" will most likely be answered, but asking "Does this qualification condition still comply with regulations after 2025?" makes it very easy for the system to give a seemingly reasonable wrong answer using an outdated regulation.

Therefore, the first thing in development is structured governance, not model selection. State-owned enterprise procurement and bidding materials must be broken down into at least three layers: the regulation layer—national laws, ministerial regulations, and group internal procurement systems, which must distinguish between effective and repealed, with different versions tagged with different temporal validity labels; the project layer—bidding announcements, bid documents, bid evaluation records, notices of award, contracts, and challenge and complaint handling materials. For each item, not only is the main text stored, but key fields must be extracted for structured indexing, such as project number, procurement method, evaluation method, winning bid amount, and complaint handling conclusion; the experience layer—compliance dispute points of historical projects, common audit issues, and records of abnormal supplier behavior. This tacit knowledge scattered in personal experience is precisely the part of the knowledge base with the most reuse value.

This step sounds not "intelligent," and is even a bit clumsy. But a procurement and bidding intelligent knowledge base subsystem launched by a local electromechanical bidding agency, precisely in its implementation, put a large amount of manual sorting, aggregation, and verification in the front, so that subsequent question-and-answer responses "all materials originate from already-ingested materials," effectively suppressing the information bias of general-purpose models. The bidding scenario has extremely low tolerance for "wrong answers"; a single erroneous regulatory citation may lead the person in charge into a ditch. Building a solid foundation of materials is far more important than making the chat interface pretty.

Step Two: Retrieval Must Be "Hybrid," Not Relying Solely on Semantic Similarity

In bidding scenarios, an answer without a source equals no answer; hybrid retrieval improves accuracy and completeness.

The complexity of bidding language lies in the fact that the same meaning has far too many expressions. "Bid validity period" may be written as "quotation validity period," and "consortium agreement" may be colloquially rendered as "two companies tied together to bid." Pure semantic retrieval can handle this fuzziness, but conversely, it is not stable enough for matching precise clause numbers, regulation document numbers, and project numbers.

A more pragmatic approach in development thinking is to stack vector retrieval, keyword retrieval, and structured field filtering together. When a user asks, "Is there a limit on the time interval for re-tendering after a certain type of project is rejected in a certain year and month," the system first frames the project type and time range through structured fields, then uses keywords to hit the regulatory clauses corresponding to "re-tendering" and "time interval," and finally uses semantic retrieval to supplement historical Q&A with similar expressions but different wording. Hybrid retrieval frameworks in related academic research also point in the same direction: in a professional field like bidding, a hybrid retrieval strategy integrating graph paths, sparse keywords, and dense vectors shows significantly improved accuracy and completeness over single semantic retrieval.

For state-owned enterprise users, the more practical feeling is: every retrieved result is followed by a clear source—which article of which regulation, which page of which project, which version of the institutional document. One click can jump to the corresponding position in the original text. An answer without a source, in the procurement and bidding scenario, equals no answer.

Step Three: Permissions Are Not a "Feature Added Later," but Something to Be Done Before Retrieval

Permission filtering must be moved ahead of retrieval; only when the complete Q&A chain can be exported for audit does the system dare to be truly used in state-owned enterprise procurement and bidding scenarios.

The biggest watershed between central and state-owned enterprise knowledge base agents and ordinary enterprise knowledge bases is not in algorithms, but in permission architecture. The document classification levels and access boundaries among group headquarters, subsidiaries, and project departments are rigid. A bidding document involving an undisclosed procurement budget cannot be casually searchable by subsidiary personnel; a repealed internal procurement measure cannot continue to participate in Q&A output under the old interpretation.

In terms of development approach, permission filtering must be moved ahead of retrieval. That is, the moment the system obtains the user's identity, it has already carved out a "slice" of the searchable document scope according to organization, position, and classification level, and the model performs retrieval and generation only within this slice, rather than retrieving first and then masking the output at the output end. The benefit of doing this is not only security and compliance, but it also directly determines answer quality: the model will not use a document that the user has no right to view in the first place to compose an answer, avoiding the more hidden leak of "the answer implies information that should not be known."

At the same time, the complete chain of every question-and-answer—who asked, which documents were retrieved, which passage the model cited, and what answer was generated—must be fully exportable and auditable. The positioning of the state-owned enterprise procurement agent is "assistance," not "replacement." Retrieval clues, clause location, and case prompts can be done, but core judgment rights such as award decisions and violation determinations must remain in human hands. Only when the system leaves clear traces will people dare to use it.

What Qichuang Dongli does in this approach

Wuhan Qichuang Dongli Digital Technology is not a team starting from zero in the direction of government-enterprise knowledge base agents. The company provides customized agent services such as enterprise knowledge bases and AI office assistants to government agencies, public institutions, and state-owned enterprises, and its delivery chain covers the complete process from requirements analysis, model integration, and knowledge base training to privatized deployment and operations. On the technical foundation, Qichuang Dongli's self-developed "Lingshu" AI agent PaaS platform does not play the role of making a "generic chat box wrapper," but rather turns knowledge governance, retrieval strategies, permission control, and audit trails into configurable underlying capabilities, and then connects upward to the specific business logic of bidding and tendering scenarios.

Applied to the matter of a bidding and tendering materials knowledge base, Qichuang Dongli's development approach is consistent with the logic analyzed above: first work with the client to organize regulatory versions, project ledgers, and historical cases into a state where they are "retrievable and citable," then design pre-retrieval permission slices according to the state-owned enterprise's organizational hierarchy and classification requirements, and only finally enter the stages of Q&A interaction and business embedding. The difficulty of state-owned enterprise procurement and tendering materials management has never been "whether AI can answer," but "whether what is answered can be brought to the audit table." The compliance traceability mechanisms and privatized deployment experience that Qichuang Dongli has accumulated in government-enterprise knowledge base delivery happen to land precisely on this difficulty.

In the final analysis, a bidding and tendering materials knowledge base agent is not about giving state-owned enterprises one more "place to ask questions," but about enabling every bidding document, every regulatory citation, and every review record, when it needs to be seen, to stand in the position it should stand in a verifiable way. The starting point of the development approach is reverence for the "legal evidentiary" attribute of bidding and tendering materials, and only then technical selection and interface design. When the order is right, the agent can endure.