Knowledge Base Search

Source: https://github.com/aipoch/medical-research-skills

Knowledge Base Search

When to Use

• You need to find specific facts, definitions, or procedures from a local knowledge base and return the exact source location.
• You must provide traceable citations (file path + page/paragraph/section) for audit, compliance, or review.
• You need to verify the original wording of a claim in the source document (quote-level validation).
• You want to compare how multiple local documents discuss the same topic and identify differences.
• You need to assemble supporting snippets for a report, FAQ, or internal knowledge response using only local materials.

Key Features

• Supports multiple retrieval approaches: direct file search, index-based search, and PageIndex-style location mapping.
• Query strategy guidance: keyword splitting, synonym expansion, and optional filters (time range, file type, tags).
• Relevance-oriented result ranking and filtering to keep the most supportive evidence first.
• Outputs verifiable hit snippets with precise citation locations (file + page/paragraph/section when available).
• Enforces local-only boundaries: searches only within authorized directories and does not modify source content.

Dependencies

• glob (>= 10.0.0): file path pattern matching
• grep (>= 3.11): in-file text searching
• Local knowledge base index files (one or more of: filename index, content index, vector index, PageIndex mapping)
• assets/hit_list_template.csv: standardized hit list output template
• Optional reference: references/guide.md (output formats, checklists, inspection points)

Example Usage

The following example demonstrates an end-to-end local search workflow and produces a CSV hit list compatible with assets/hit_list_template.csv.

Inputs

• Knowledge base root: ./kb/
• Query: How do we rotate API keys?
• Filters: file types md,pdf, time range 2024-01-01..2026-12-31

Steps

1. Confirm index and scope

   • Ensure the search scope is limited to authorized paths (e.g., ./kb/).
   • Identify available indices:
     • filename/content index (fast keyword search)
     • vector index (semantic retrieval)
     • PageIndex mapping (page/paragraph location resolution)

2. Build the query

   • Keywords: rotate, API key, key rotation
   • Synonyms/variants: credential rotation, token rotation, regenerate key
   • Filters:
     • file type: *.md, *.pdf
     • time range: 2024-01-01..2026-12-31 (if metadata exists)

3. Execute search (local-only)

   • Path discovery (example):
     • glob("./kb/**/*.md")
     • glob("./kb/**/*.pdf")
   • Content search (example):
     • grep -RIn "API key\|key rotation\|rotate" ./kb/

4. Filter and rank results

   • Keep hits that directly answer the question (procedure, policy, steps, constraints).
   • Rank by:
     • term proximity (e.g., “rotate” near “API key”)
     • section relevance (e.g., “Security”, “Credentials”, “Operations”)
     • coverage (hits that include prerequisites + steps + verification)

5. Output citations and hit list

   • For each hit, output:
     • file_path
     • location (page number for PDFs; heading/paragraph index for Markdown; PageIndex if available)
     • snippet (verbatim excerpt supporting the conclusion)
     • notes (why it is relevant; any assumptions)
   • Save as hit_list.csv using assets/hit_list_template.csv columns.

Example Output (CSV rows)

file_path,location,snippet,relevance_score,notes
kb/security/credential_policy.pdf,page 12,"API keys must be rotated every 90 days... Rotation requires...",0.92,"Direct policy + rotation interval + procedure reference."
kb/runbooks/api_key_rotation.md,section 'Procedure' ¶3,"To rotate an API key: (1) create a new key... (2) update services... (3) revoke old key...",0.89,"Step-by-step operational runbook."
kb/audit/controls.md,heading 'Key Management' ¶2,"Evidence of rotation includes change tickets and key revocation logs...",0.81,"Provides verification/evidence requirements."

Implementation Details

Retrieval Workflow

1. Index confirmation

   • Determine knowledge base root paths and last update time (if available).
   • Detect which indices exist:
     • filename index: quick narrowing by file names
     • content index: inverted index / grep-like scanning
     • vector index: semantic similarity retrieval
     • PageIndex: mapping from document offsets to page/paragraph identifiers

2. Query strategy

   • Tokenize the question into:
     • core entities (e.g., “API key”)
     • actions (e.g., “rotate”, “revoke”, “regenerate”)
     • constraints (e.g., “every 90 days”, “approval required”)
   • Expand with synonyms and variants.
   • Apply filters when metadata exists:
     • time range
     • file type
     • tags/collections

3. Result filtering and ranking

   • Remove low-signal hits (navigation, boilerplate, unrelated mentions).
   • Rank by a weighted score (example):
     • Keyword match (exact phrase > partial): 0.45
     • Proximity (terms close together): 0.20
     • Section importance (titles like “Procedure/Policy”): 0.20
     • Coverage (answers include steps + constraints + verification): 0.15
   • Keep the original text snippet verbatim for verification.

4. Citation and location resolution

   • Markdown/text:
     • use heading + paragraph index (or line range) as the primary locator
   • PDF:
     • use page number; optionally include bounding text around the hit
   • PageIndex (if present):
     • map internal offsets to stable page/paragraph identifiers

Constraints and Limitations

• Search only within user-authorized local directories.
• Do not modify source documents.
• Do not execute scripts or arbitrary code.
• Do not access network resources or external APIs.
• If indices are missing/corrupted, fall back to direct file scanning; if scanning is not possible, report the limitation and required remediation (re-indexing).