AI对话知识管理工具

pip install project-tessera
tessera setup
# Done. Claude Desktop now has persistent memory + document search.

    User asks: "What did we decide about the database?"
                            |
                            v
                +-----------------------+
                |    Query Processing   |
                |  Multi-angle decomp   |    "database decision"
                |  (2-4 perspectives)   |    "database", "decision"
                +-----------------------+    "decision about database"
                            |
              +-------------+-------------+
              |                           |
              v                           v
    +------------------+        +------------------+
    |  Vector Search   |        |  Keyword Search  |
    |  (LanceDB)       |        |  (FTS index)     |
    |  384-dim MiniLM  |        |  BM25 scoring    |
    +------------------+        +------------------+
              |                           |
              +-------------+-------------+
                            |
                            v
                +-----------------------+
                |      Reranking        |
                |  70% semantic weight  |    LinearCombinationReranker
                |  30% keyword weight   |    + version-aware scoring
                +-----------------------+
                            |
                            v
                +-----------------------+
                |   Result Assembly     |
                |  Dedup (content hash) |    2-pass deduplication
                |  Verdict labels       |    found / weak / none
                |  Cache (60s TTL)      |
                +-----------------------+
                            |
                            v
                    Top-K results with
                    confidence scores

    Documents: .md .pdf .docx .xlsx .py .ts .go ...  (40+ types)
                            |
                            v
                +-----------------------+
                |   File Type Router    |
                |  Markdown, CSV, XLSX  |    Type-specific parsers
                |  Code, PDF, Images    |    with metadata extraction
                +-----------------------+
                            |
                            v
                +-----------------------+
                |   Chunking Engine     |
                |  1024 tokens/chunk    |    Sentence-boundary aware
                |  100 token overlap    |    Heading-preserving
                +-----------------------+
                            |
                            v
                +-----------------------+
                |   Local Embedding     |
                |  fastembed/ONNX       |    paraphrase-multilingual
                |  384 dimensions       |    MiniLM-L12-v2
                |  No API calls         |    101 languages
                +-----------------------+
                            |
              +-------------+-------------+
              |                           |
              v                           v
    +------------------+        +------------------+
    |    LanceDB       |        |     SQLite       |
    |  Vector storage  |        |  File metadata   |
    |  Columnar format |        |  Search analytics|
    |  Zero-config     |        |  Interaction log |
    +------------------+        +------------------+

                    +--------------------------------------------+
                    |              src/core.py                    |
                    |         58 orchestration functions          |
                    |   69 specialized modules, 31k LOC           |
                    +--------------------------------------------+
                     /                |                \
    +---------------+  +-------------------+  +--------------+
    | MCP Server    |  | HTTP API Server   |  | CLI          |
    | Claude Desktop|  | FastAPI + Swagger |  | 11 commands  |
    | 58 tools      |  | 58 endpoints      |  | setup, sync  |
    | stdio         |  | port 8394         |  | ingest, api  |
    +---------------+  +-------------------+  +--------------+
           |                    |                     |
           v                    v                     v
    +------------------------------------------------------------+
    |                    Storage Layer                            |
    |  LanceDB         SQLite           Filesystem               |
    |  (vectors)       (metadata,       (memories as .md,        |
    |                   analytics,       encrypted with           |
    |                   interactions)    AES-256-CBC)             |
    |                                                            |
    |  fastembed/ONNX: local embedding, no API keys              |
    |  101 languages, 384-dim vectors, ~220MB model              |
    +------------------------------------------------------------+

pip install project-tessera

uvx --from project-tessera tessera setup

tessera setup

tessera api                     # Start REST API on localhost:8394
ngrok http 8394                 # Expose to the internet
# Then create a Custom GPT with the Actions spec from /chatgpt-actions/openapi.json

# Via MCP (Claude)
"Remember that we chose PostgreSQL for the production database"

# Via HTTP API (scripts, local LLMs, ChatGPT via tunnel)
curl -X POST http://127.0.0.1:8394/remember \
  -H "Content-Type: application/json" \
  -d '{"content": "Use PostgreSQL for production", "tags": ["db", "architecture"]}'

CONTRADICTION (HIGH severity):
  "We decided to use PostgreSQL" (2026-03-01)
  vs
  "Switched to MongoDB for the main database" (2026-03-10)

  The newer memory (2026-03-10) likely reflects the current state.

# 1. Start Tessera API + tunnel
tessera api
ngrok http 8394   # or: cloudflared tunnel --url http://localhost:8394

# 2. Get the OpenAPI spec for your Custom GPT
curl https://your-tunnel-url/chatgpt-actions/openapi.json?server_url=https://your-tunnel-url

# 3. Get the GPT instruction template
curl https://your-tunnel-url/chatgpt-actions/instructions

curl -X POST http://127.0.0.1:8394/import-conversations \
  -H "Content-Type: application/json" \
  -d '{"data": "<ChatGPT export JSON>", "source": "chatgpt"}'

curl http://127.0.0.1:8394/export?format=obsidian

# workspace.yaml
hooks:
  on_memory_created:
    - script: ./notify-slack.sh
  on_contradiction_found:
    - script: ./alert.py

pip install project-tessera[api]
tessera api  # http://127.0.0.1:8394

# Search documents
curl -X POST http://127.0.0.1:8394/search \
  -H "Content-Type: application/json" \
  -d '{"query": "database architecture", "top_k": 5}'

# Save a memory
curl -X POST http://127.0.0.1:8394/remember \
  -H "Content-Type: application/json" \
  -d '{"content": "Use PostgreSQL for production", "tags": ["db"]}'

# Export memories
curl http://127.0.0.1:8394/export-for-ai?target=chatgpt

# Batch (multiple operations, single request)
curl -X POST http://127.0.0.1:8394/batch \
  -H "Content-Type: application/json" \
  -d '{"operations": [{"method": "search", "params": {"query": "test"}}, {"method": "knowledge_stats"}]}'

tessera setup          # One-command setup (config + model download + Claude Desktop)
tessera init           # Interactive setup
tessera ingest         # Index all document sources
tessera sync           # Re-index changed files only
tessera serve          # Start MCP server (stdio)
tessera api            # Start HTTP API server (port 8394)
tessera migrate        # Upgrade data schema
tessera check          # Workspace health diagnostics
tessera status         # Project status summary
tessera install-mcp    # Configure Claude Desktop
tessera version        # Show version

{
  "mcpServers": {
    "tessera": {
      "command": "uvx",
      "args": ["--from", "project-tessera", "tessera-mcp"]
    }
  }
}

{
  "mcpServers": {
    "tessera": {
      "command": "tessera-mcp"
    }
  }
}

workspace:
  root: /Users/you/Documents
  name: my-workspace

sources:
  - path: .
    type: document

search:
  reranker_weight: 0.7     # Semantic vs keyword balance (0.0 = keyword only, 1.0 = vector only)
  max_top_k: 50            # Max results per search

ingestion:
  chunk_size: 1024         # Tokens per chunk
  chunk_overlap: 100       # Overlap between chunks

hooks:                      # Optional plugin hooks
  on_memory_created:
    - script: ./my-hook.sh