Vector-store content retrieval

Use vector-store content retrieval when an agent needs to ask conversational questions of a full source-file package instead of relying only on record metadata or short snippets.

This is the agentic source-package review path for solicitation packages, PWS/SOW files, amendments, Q&A, pricing instructions, exhibits, and workspace files. Find the source package with Search_Government_Files, choose the files to stage, then retrieve source-grounded passages without manually downloading each attachment. Analyst and Analyst Astra create a store on demand; external MCP clients can start with Add_To_Vector_Store.

This pattern is for government files and user files staged through Add_To_Vector_Store. It is not a GovTribe documentation search, a general GovTribe knowledge base, or a full-document export.

Analyst and Analyst Astra workflow

An Analyst conversation starts without a vector store. When full-text retrieval is needed, call the root Create_Vector_Store function with {"action":"ensure","request_id":null}. Poll it with {"action":"status","request_id":"<REQUEST_ID>"} until the store is ready. Stop polling and follow the returned guidance if creation fails or times out. The returned govtribe_vector_store_id identifies the current store; the request_id stays the same if an expired store is replaced. Add_To_Vector_Store does not create the initial Analyst store.

The agent then chooses an opportunity or exact government/user files and calls Add_To_Vector_Store with those items and the current store ID. Follow Add's existing polling instructions until the selected files are ready. Call Create_Vector_Store with action: status once more using the original request ID. When it reports file_search_available: true, hosted file_search can retrieve the indexed text in the same conversation turn. A ready but empty store does not make file_search available.

The agent decides which files to add. If a prior store expired, status lists its previously_attached_files from the stored file relationships, including each file's type, ID, and name. Review that list, choose the files still needed, and pass those exact type/ID pairs to Add_To_Vector_Store with the current store ID. Earlier files are not copied automatically. Create_Vector_Store is available to Analyst and Analyst Astra, not external MCP clients. External clients continue with the Add → Search_Vector_Store workflow below.

Vector-store retrieval does not cover every file type in a source package. If Add_To_Vector_Store reports skipped or unsupported spreadsheet-like files, such as pricing schedules, workbooks, CSV/TSV files, CLIN tables, budget templates, or staffing matrices, use Add_To_Container and hosted shell extraction before making complete package-level claims.

Add_To_Container can also stage an existing accessible workspace User File without a new conversation upload, or stage one typed-search page as JSON for Python, chart, spreadsheet, document, or dataset-join work. Use Use search results and user files in the hosted shell to choose that path. Continue using vector stores when the task is semantic retrieval from supported file text.

Solicitation package review workflow

Use this workflow when the user asks for a conversational readout of an entire solicitation package, not just a list of files.

Choose the starting point

When to use vector-store retrieval

Use this pattern when the user needs semantic retrieval from file text, such as:

• requirements, clauses, instructions, or risks buried inside a solicitation attachment
• exact supporting evidence from a user-uploaded file
• concept-level matching across long file text where keyword metadata search is too thin
• a focused answer that depends on retrieved chunks from one or more staged files

Use normal search tools first when file metadata, record fields, or content_snippet values answer the question.

Do not require content_snippet or parent_record to continue a source-package review workflow. Some clients or search paths may omit those fields even when the file can still be staged and searched. To link a file back to a parent, prefer the ID you filtered on, the resolved GovTribe URL, file name, and file-source metadata returned by the tool.

Decide whether to stage files

Prefer the least expensive path that can answer the question.

Before staging files

Resolve the file or file-bearing entity before calling Add_To_Vector_Store.

Supported staged item types currently include file records and selected file-bearing entities such as government_file, user_file, pursuit, federal_contract_opportunity, federal_contract_idv, and federal_grant_opportunity.

Add files to a vector store

Call Add_To_Vector_Store with the resolved items. In Analyst or Analyst Astra, first ensure the store with Create_Vector_Store and include its current govtribe_vector_store_id in Add. For an external MCP client, omit the ID to reuse the current conversation store or create a standalone store when there is no conversation; include an existing ID to append files. An expired selected store is replaced.

Tool: Add_To_Vector_Store

{
  "govtribe_vector_store_id": "<VECTOR_STORE_ID>",
  "items": [
    {
      "govtribe_type": "government_file",
      "govtribe_id": "<GOVERNMENT_FILE_ID>"
    }
  ]
}

Capture the returned govtribe_vector_store_id. Later steps use that ID for readiness checks and retrieval. If an external MCP client's first call omitted the ID, add the returned ID to follow-up Add_To_Vector_Store status checks while keeping the same source items; otherwise it can create a new vector store instead of polling the existing operation.

Expired stores

When the selected vector store has expired, Add_To_Vector_Store creates a replacement and adds only files requested by the affected add operation. Earlier files are not carried over automatically. The response message discloses replacement on the initial result and later polls. Use the returned new govtribe_vector_store_id for polling and subsequent calls, and explicitly add any earlier files you still need. Wait for successful processing before searching; pending or failed files are not searchable.

Interpret Add_To_Vector_Store output

Use Add to vector store MCP response as the source of truth for response shape, text-only client behavior, retry instructions, ready-for-search guidance, and failure handling.

Poll readiness

Follow the Add-to-vector-store response status until the requested files are ready.

For complete package analysis, wait until Add_To_Vector_Store reports that the requested files are ready. In Analyst or Analyst Astra, refresh Create status before using hosted file_search. For external MCP clients, a premature Search_Vector_Store call can return a not-ready message. Some clients may return available chunks while other requested files are still pending; treat those results as partial and keep polling before making final package-level claims.

Search staged file content

External MCP clients call Search_Vector_Store after the vector store is ready. Analyst and Analyst Astra use hosted file_search after Create status reports file_search_available: true. Keep max_num_results as small as the question allows when using Search_Vector_Store.

Focused retrieval:

Tool: Search_Vector_Store

{
  "query": "List mandatory deliverables and submission format instructions.",
  "govtribe_vector_store_id": "<VECTOR_STORE_ID>",
  "max_num_results": 5,
  "rewrite_query": false
}

Conceptual retrieval:

Tool: Search_Vector_Store

{
  "query": "Summarize cybersecurity incident response expectations and reporting deadlines.",
  "govtribe_vector_store_id": "<VECTOR_STORE_ID>",
  "max_num_results": 10,
  "rewrite_query": true
}

Treat returned chunks as semantic evidence. They can support an answer, but they are not a full-document export.

Query construction

For solicitation package reviews, run a small set of focused searches instead of one broad search:

• scope, deliverables, and performance requirements
• submission instructions, volume structure, page limits, required forms, and due-date rules
• evaluation factors, pass/fail requirements, and scoring criteria
• pricing instructions, option-year assumptions, and contract type
• eligibility, set-aside, facility, clearance, cybersecurity, and compliance constraints
• amendments, Q&A, or attachments that change instructions or deadlines

External MCP examples

External MCP clients can use this sequence when full file text is needed:

1. Resolve the target record. Use Search_GovTribe or the relevant typed search tool to identify the opportunity or file-bearing record.
2. Find the attached files with Search_Government_Files or Search_User_Files, and choose the smallest set that answers the question.
3. Call Add_To_Vector_Store with the chosen items. Omit govtribe_vector_store_id for a new standalone store, or pass an existing ID to add to that store.
4. While Add reports in_progress, call it again later with the same items and its returned govtribe_vector_store_id. Review skipped and failed files before treating a package as complete.
5. When the requested files are ready, call Search_Vector_Store with the returned store ID and focused source questions. Treat chunks as evidence, and disclose any missing attachment or incomplete coverage.

Find candidate government files before staging a package:

Tool: Search_Government_Files

{
  "query": "PWS SOW Section L Section M proposal instructions evaluation factors",
  "federal_contract_opportunity_ids": ["<FEDERAL_CONTRACT_OPPORTUNITY_ID>"],
  "fields_to_return": [
    "govtribe_id",
    "govtribe_url",
    "name",
    "govtribe_ai_summary",
    "posted_date",
    "download_url"
  ],
  "per_page": 10
}

Stage an entire opportunity package when the user asks for a full solicitation package review:

Tool: Add_To_Vector_Store

{
  "items": [
    {
      "govtribe_type": "federal_contract_opportunity",
      "govtribe_id": "<FEDERAL_CONTRACT_OPPORTUNITY_ID>"
    }
  ]
}

Retrieve source evidence for submission and evaluation rules:

Tool: Search_Vector_Store

{
  "query": "Extract proposal volume requirements, page limits, required forms, submission method, and evaluation factors.",
  "govtribe_vector_store_id": "<VECTOR_STORE_ID>",
  "max_num_results": 10,
  "rewrite_query": false
}

Stage one file when the question is limited to a specific attachment:

Tool: Add_To_Vector_Store

{
  "items": [
    {
      "govtribe_type": "government_file",
      "govtribe_id": "<GOVERNMENT_FILE_ID>"
    }
  ]
}

Retrieve clause-level evidence from that file:

Tool: Search_Vector_Store

{
  "query": "What cybersecurity reporting, incident response, and access control requirements appear in this attachment?",
  "govtribe_vector_store_id": "<VECTOR_STORE_ID>",
  "max_num_results": 6,
  "rewrite_query": true
}

When not to use it

Do not stage files just because vector retrieval is available. Prefer the normal MCP search path when:

• the question is about record metadata, not file text
• content_snippet already answers the question
• the user needs a list of files, not semantic chunks from file contents
• the selected govtribe_type is not supported by Add_To_Vector_Store

Related articles

• Add to vector store MCP tool: Review supported item types and staging arguments.
• Search vector store MCP tool: Review semantic retrieval arguments.
• Search vector store MCP response: Interpret returned structured results, snippets, and readiness messages.
• Search government files: Search government-file metadata and snippets before staging full text.
• Search user files: Search team-uploaded file metadata and snippets before staging full text.
• Proposal Workflows with MCP: Use the generated Solicitation Package Review prompt as a copy-ready workflow in connected MCP clients.